Artificial Intelligence – Julian Green, Michael Jason Grundmann, Sylvia Joan Smullin, Joseph Pieter Stefanus van Grieken, X Development LLC

Abstract for “Dynamic Traffic Control”

“Some implementations include receiving camera data from a processing device that controls a traffic signal at an intersection. The processing devices are located near the intersection. They then use one or several local machine-learning models to identify objects at intersections and their paths based on these images. Finally, they transmit traffic data from remote traffic planning systems over a network. This remote instruction is determined based (i) on remote traffic planning software and (iii) on local instructions generated by processing devices.

Background for “Dynamic Traffic Control”

“This specification generally refers to traffic indicators. Traffic indicators are used to alert vehicles, other objects and travellers along a road. Traffic indicators can be traffic signs, traffic lights, or audio cues to cross. Local controllers control traffic lights by instructing them to change their phase (e.g. from red to green). Traffic congestion is affected by the timing of traffic lights at adjacent intersections. It also affects drivers’ travel time and wait times.

A traffic light control system can regulate traffic lights’ phases at intersections or multiple intersections to maintain smooth and safe traffic flow. Traffic light control systems can identify the appropriate phases of traffic lights using local traffic data at the intersections where they are being controlled and global instructions generated by analyzing traffic data at many other intersections.

“For example, traffic light phases can be coordinated to cut travel time by scheduling traffic light phases so that cars passing through intersections come across a series of green lights or a dynamic green wave. To protect pedestrians crossing intersections, the phases of traffic lights can also be controlled by moving the green light phase at intersections that direct traffic in the pedestrian’s directions. To protect pedestrians, bicyclists and other nonvehicular travellers, it is possible to control lights that inform pedestrians and bicyclists when crossing the intersection is safe. The lights can be programmed to give non-vehicular travelers a lead time (e.g. a headstart) or balance the timing with vehicles who would be passing the non-vehicular traveller.

The traffic light control system controls the phases of traffic lights and can reduce unsafe driving habits. Drivers can be encouraged to follow posted speed limits by having traffic lights at nearby intersections coordinate to change phase. Coordinated traffic lights reduce congestion and reduce the need for roads to be expanded to handle disproportionately high volumes of traffic during peak travel times. They also reduce fuel consumption and vehicle pollution. The traffic light control system reduces frustration among drivers by reducing travel times.

“The combination global instructions and local data allow the traffic control system learn from traffic situations at multiple intersections, while still being flexible based on local conditions. Global instructions might suggest that traffic can be reduced by reducing red light phases at traffic lights. Global instructions can be helpful for many intersections to alleviate congestion. Local data may indicate that a bicyclist crosses an intersection. If the traffic light changes to a green phase, the bicyclist won’t be able finish crossing the intersection. Traffic control systems may alter the instructions to the traffic lights to prolong the red light phase until the bicyclist crosses the intersection. Separate lights such as the walk lights for pedestrians are sometimes used to protect non-vehicular travelers.

“Accordingly to an innovative aspect of this specification, a method of controlling traffic lights is provided. In this system, sensor data providing images of traffic intersections is collected by cameras or other sensors located near the intersection. The sensor data may be collected by sensors at an intersection and used to identify objects using local machine learning models. The processing devices might determine the object’s position relative to the intersection, speed, and direction of travel. Traffic planning systems located far from intersections may be able to receive traffic data including identified objects from local machine learning models. This information can be transmitted over a network. The instructions may be generated by the remote planning system and sent to the processing devices controlling the traffic light. The remote planning system may send instructions to the processing devices controlling the intersection’s traffic lights. These instructions can be used to generate local control instructions. The remote planning system instructions and local instructions may be used by a traffic light controller to generate and give a control instruction.

One innovative aspect of this specification is that it includes receiving camera data from intersections, which can then be used to control traffic signals at intersections. The one or two processing devices are located near the intersection. They also use local machine-learning models to identify objects at intersections and their paths. A remote instruction for traffic signal control at intersections is sent to remote traffic planning systems over a network. This remote instruction is determined based (i) on remote instructions from remote traffic planning systems and (ii) on local instructions generated by one or several processing device.

Implementations can include any or all of the following features. The method could include, for example, accessing safety rules that indicate requirements for traffic signals, modifiying the remote instruction based upon the safety rules, and then providing the modified instruction to traffic signal by one or more processing devices. The method may include determining a condition at an intersection using one or more of the local machine learning models, altering the remote instruction based upon the local instruction and providing the control instruction to traffic signal.

“In some cases, the method involves determining by one or more processors and using one or two local machine learning models, and further based upon the camera data and phase information for the traffic signals indicating a traffic signal’s current phase, a classification of a condition at intersection, and generating by one or more processors a local instruction based the classification and overriding, depending on the remote instruction. The method may include determining the number of objects traversing a path through an intersection using one or more processing units. This includes counting the objects of a specific type that traverse the path. Claim 5, wherein the method includes providing traffic data that indicates the speed at which the counted objects are moving.

“In some cases, providing traffic data may include providing one or more measures that indicate relative numbers of vehicles using different routes through the intersection. Some examples show reinforcement learning, where traffic data from nearby intersections is received and taken into consideration by the remote planning system.

“In some cases, dynamically altering the traffic signal by one or more processing device and based upon the traffic instruction is part of the method. The one or more processing units are usually located at the intersection in some cases. The method may include determining the location of the one or more processing devices based at least partially on traffic data. The method may include providing traffic data from local machine learning models over a network to a remote traffic planning software via the one or several processing devices periodically.

“In some cases, the method may include receiving custom machine learning model parameters for a local machine learning program. The method may include determining the control instructions to the traffic signal at an intersection using a real-time operational system. This is based on safety rules. The method may include using one or more local machine-learning models to classify an intersection condition or assess the speed of an identified object. The method may include storing parameters to generate local control instructions by one or several processors in a database. Accessing the database to retrieve these parameters by one or two processors allows them access the parameters. Finally, they can generate a local control instruction based on the outputs of one or multiple local machine learning models.

“The subject matter in this specification may be implemented in specific embodiments to achieve one or more of these advantages. This system is flexible and can be used with sensor systems such a lidar radar, range imaging microwave sensors, inductive loops black and white cameras, color cameras like red green (RB) cameras, infrared cameras, hyper-spectral camera, multi-spectral cameras (IR) cameras, directional sensors. Other types of data can be collected in some implementations. Data associated with traffic indicators may include data about railroad crossings and traffic signals, such as traffic lights, railroad crossings, flags, signs, road barriers, flags and traffic signs. Data may also be collected. You may also use other sensors or combinations of sensors. This system decreases the wait time for vehicles at intersections by optimizing and controlling traffic lights to control traffic flow. Safety and traffic throughput are also improved. It can also reduce the number and frequency of stops that any traveller (vehicle or otherwise) must make at traffic lights. The system can analyze traffic data to better understand traffic flows for specific modes by analysing them. The system can help you understand the traffic patterns of specific modes. For instance, pedestrians might have a different flow pattern to buses than they do to passengers cars. Monitoring traffic behavior can help to improve traffic safety at intersections. The system can identify factors that could lead to accidents or near-collisions. Signal timing can be used to give priority to certain travelers in some cases. The system can allow pedestrians to cross the road before bicyclists or other non-vehicular travellers. Additional lights such as a warning lamp can be used in certain cases. A warning light can be added to a traffic signal, for example, to warn drivers when a collision is imminent.

“The accompanying drawings and description below detail one or more implementations for the subject matter. The claims, drawings, and description will reveal other potential features, aspects, or advantages to the subject matter.

“Some traffic control systems include multiple levels of machine-learning processing. A remote system (e.g., server system or cloud-based) may use reinforcement learning for traffic control coordination among multiple traffic lights within an area, such as a city. Local processing systems that run other machine learning models may also be distributed in the area. For example, a local system can be installed at each intersection. The local systems monitor intersection conditions and send traffic data to remote systems. Remote systems use its learning models to generate instructions which are then sent to local systems. This arrangement allows traffic lights to be optimized or coordinated at the network level. However, it also gives local intersections the ability to adapt to local conditions using local machine learning models. The remote system’s instructions may improve traffic throughput and efficiency over large areas. However, each local processing unit might be able override these instructions to address local conditions, such as elderly people crossing the street or emergency vehicles approaching.

“In some cases, a traffic control systems controls a traffic light at an intersection. A local processing system runs a real time operating system (RTOS) and controls each traffic light. The local processing system may be located near the intersection where the traffic lights are located. Other modules within the local processing network may send instructions to the RTOS, including a local planning module, which is part of the local system, or a remote planning system, which sends data over the internet. Local machine learning models may be used by the local planning module to identify intersection objects and generate traffic control instructions based on the local conditions. Remote machine learning models may be used by the remote planning system to generate traffic light control instructions based on conditions at different intersections. The remote planning system could receive traffic data from multiple intersections and use reinforcement learning to improve traffic flow over a large area with many intersections.

“FIG. “FIG. A traffic intersection is generally composed of one or more traffic signals and connecting two or more roads. The techniques described in this article can be used to control any traffic signal. A traffic intersection is a three-way intersection, which connects one end of a road with another length. It can have three traffic lights, one for each segment of road or arm. It can also have arms extending from it. Traffic intersections may have more or less arms depending on how they are implemented. This particular example shows a four-way intersection 100 that connects two roads. Each arm of the intersection is represented by a separate traffic light, 102 a-d.

In some cases, different sets of lights may be included at each intersection’s arm. An intersection might have a number of lights, such as one for bicyclists and another for passengers. Traffic lights may include audio cues as an alternative or addition to visual cues. Traffic lights might include a pedestrian light and speaker output to indicate that it is safe to cross. Traffic lights may include accessibility features such as vibration modules or other similar technologies.

“In this example, several sensors, including camera104 a?d and other sensors106 a?d, are shown as being proximate each traffic light (102 a?d). Cameras 104 capture information such as images or other sensor data for each arm. To capture information about an intersection, multiple cameras or sensors can be placed in different locations, such as at traffic lights that are not immediately adjacent to them. There may be several sensors per traffic light in some cases. A single camera or sensor may capture information about multiple lanes, arms or areas of a roadway. Multiple sensors are unnecessary. There may be multiple cameras for each traffic light in some cases.

“The traffic data captured at each traffic light 102 a-d can be transmitted to a local control network 110. In some cases, each intersection is controlled by a local control system. One or more intersections may be controlled by a local control system. A local control system could control a group of intersections that are close by. A local control system could control multiple intersections that are located far from one another. A local control system may be responsible for controlling a specific section of an intersection. A local control system might control one of the intersection’s four traffic lights, while another local control system could control the intersection’s other three traffic lights. Different local control systems could control different numbers.

“In some cases, the local control systems 110 may include one or more local machine-learning models 112. To process traffic data, the local machine learning models 112 will be used. The local machine learning models 112 may store data in a local control system’s memory. The local control system 110 may store metadata that is determined by local machine learning models 112 rather than sensor data from cameras 104 or sensors 106. The local control system 110 might store, for example, the number and speed of vehicles turning right at the intersection’s western arm, as well as the number and speed of vehicles passing through the intersection within the last five minutes.

“The local machine-learning models 112 could accept inputs from cameras and/or sensors. Local machine learning models 112 can use any number of models including decision trees, logistic regression models and neural networks. They also may use classifiers, support vector models, inductive logic programming, ensembles (e.g. using techniques like bagging, boosting or random forests). Bayesian networks, genetic algorithms, Bayesian network, etc. and can be trained using many approaches such as deep learning. Perceptrons, association rules. Inductive logic. Clustering. Maximum entropy classification. Learning classification. Some local machine learning models 112 might use supervised learning. Some local machine learning models 112 may use unsupervised learning in some cases.

“Some examples show how local machine learning models 112 can be used to identify objects at an intersection based upon traffic data. The local machine learning models 112 can also be trained to recognize specific types of objects (e.g., vehicles, busses, trucks, etc.). ), bicycles, pedestrians, debris, etc. The local machine learning model 112 can identify specific details such as make, model, license plate number and other information. characteristics of a vehicle, such as the make, model, license plate number and pedestrian characteristics. It can also identify the number of horses pulling a carriage. It can also determine the size of any debris or obstructions. The local machine learning models 112 can also be trained to recognize traffic scenes such as accidents, construction, pedestrians waiting to cross, emergency vehicles approaching, and so on. The local machine learning models 112 are able to recognize scenes and objects in certain situations. Both within the intersection and adjacent streets can be used to train the local machine learning models 112 to recognize objects, scenes, etc. The local machine learning models 112 can also be used to detect an emergency vehicle coming from an adjacent street. The local machine learning model 112 can be used to determine the speed, direction and acceleration of vehicles that pass through the intersection.

“The local machine-learning models 112 can determine the number of vehicles that pass through an intersection from a particular path. The local machine learning models 112 can identify that five vehicles travel from the western segment of intersection and that three are traveling from the southern segment. No vehicles are coming from either the eastern or northern segments of intersection. Local machine learning models 112 can calculate the speed, direction, acceleration, and acceleration of vehicles that pass through the intersection.

“The local machine-learning models 112 are able to identify traffic patterns such as the formation of queues. The local machine learning model 112 can, for example, determine that there is a lot of traffic because there are many vehicles waiting to get in the intersection. The local machine learning models 112 could also use camera data to determine if there is a large queue of vehicles waiting for exit. Sometimes, an intersection might be blocked so that no traffic is moving through it, even though the traffic lights are green. This could happen when traffic is blocked ahead of the intersection so that cars cannot pass through it. These queue data can be stored and generated to help determine if the queue conditions have cleared. Some examples show that queue data can be used to input the local learning models 112 in order to update control instructions or improve pattern recognition.

“The local machine-learning models 112 can identify the paths taken by identified vehicles. The identified paths can be described as origin-destination pair, which indicates the origin and destination of each vehicle. Some examples show the origin-destination pair as vectors. You can organize the origin-destination pairs in many ways, including lists and tables.

“The local control systems 110 generate local control instructions using traffic data collected by cameras at intersections. Local control system 110 generates local control instructions using local machine learning models 112. The traffic data may be used to customize the local control instructions for each traffic light. The local control system might receive traffic data that indicates that very few vehicles are traveling from the intersection’s northern or southern roads segments, but many are using the eastern and western road segments. The local control system 110 might generate an instruction to increase the green phase at the western and eastern segments of an intersection. It may also generate an instruction to shorten the green phase at the northern or southern segments of an intersection. The local control instructions may be generalized to the intersections or portions of intersections for which the local control system is responsible. The local control system 110 might generate instructions to reduce the yellow-to red phase transition time.

“The traffic data generated from the local machine-learning models 112 are transmitted to a remote traffic management system 122. The remote planning system 122 is connected to the local control system 110 via a network 120.

Remote planning system 122 may include separate remote machine-learning models 124. Remote planning system 122 can receive data from multiple intersections. Remote machine learning models (124) use traffic data from multiple intersections to determine global traffic control directions. Remote machine learning models 124 might use traffic data from multiple intersections to determine that the yellow-to red light transition at each intersection within the district needs to be increased by two seconds.

Remote machine learning models 124 can accept sensor data from cameras and/or sensors. Local machine learning models 112 can use any number of models including decision trees, logistic regression models and neural networks. They also may use classifiers, support vector models, inductive logic programming, ensembles (e.g. using techniques like bagging, boosting or random forests). You can train them using many methods, including deep learning, perceptrons and association rules. They can also be trained with inductive logic, clustering or maximum entropy classification. The remote machine learning models may employ supervised learning in some cases. Remote machine learning models 124 may use unsupervised learning in some cases.

“In some cases, remote machine learning models (124) uses reinforcement learning to decide which actions to take in particular traffic situations. The remote machine learning models may also use Markov decision processes. Remote machine learning models 124 could use, for example, the criterion or optimality, brute force or direct policy search or value functions including Monte Carlo or temporal differences methods.

“The remote planning software 122 can use the global traffic control instruction to generate control instructions that are transmitted to the local control system 110. The control instructions may be identical to the global traffic control instruction in some cases. The remote planning system 122 control instructions can be customized to suit each traffic light that the local control system 110 controls.

“The remote control system 122 transmits remote control instructions through the network 120 to the local control systems 110. The remote control system 122 may be used to generate the local control instructions 110. Sometimes, local control instructions can override or modify remote control instructions. If the remote control instructions reduce the length of the green-light phase but the intersection the local control 110 is monitoring is experiencing delays, the local system could instead generate local control instruction that increase the green-light phase at the intersection.

“The local control system contains an RTOS 130 which controls the traffic lights 102 a?d. Based on the remote control instruction and the local control instruction and traffic data 100 from the cameras 104 a?d and sensors 106 a?d, the RTOS 130 generates a traffic light control instruction for each traffic light 102 a?d.

“FIG. “FIG. The control system 200 comprises a local control network 210, a network 220 and remote planning system 223. A traffic light 230 is also included.

“The local control systems 210 could be the same as those described in FIG. 1. Local control system 210 contains a local planning module 212 and a communication module 214. A local machine learning model 216 is also included. The sensor processing module 218 and an RTOS 232 are also part of the local control system 210.

“The local planning module (212) receives traffic data via the local machine-learning model 216. The local planning module 212 receives traffic data and generates control instructions for traffic lights. The local planning module 212 might receive traffic data that indicates that no vehicles are present at the intersection’s western, eastern, and southern roads segments. The local planning module 212 may extend the green phase of the traffic light for northern road segments of the intersection.

“The communication module 214, which is also known as the communication module 220, allows for communication over a network (e.g. the network 220-2) and with traffic lights (e.g. the traffic light 233. The communication module 214, which can be wirelessly connected, allows for wireless communication over the network 220. The communication module 214 could be used to transmit data over wireless channels and wireless voice channels. The communication module 214 can transmit traffic data via a wireless data channel. One or more LTE modules, a GSM modul, a radio module, a cellular transmission module or any other type of module that can exchange communications in one the following formats: LTE or GPRS or CDMA, EDGE, EDGE, EVDO, UMTS or IP.

“The communication module (214) may also be a wired module that allows for communication over the network 220 via a wired connection. The communication module could be, for example, a modem or a network interface card. A communication module could be an Ethernet network card that allows the monitoring control unit 110 and the monitoring system 110 to communicate over local area networks or the Internet. The communication module 214 can be set up to communicate with traffic lights specific protocols.

“The local machine-learning model 216 receives data directly from cameras located at intersections or intersections where the local control system210 is responsible. The local machine learning model (216) may include multiple machine learning models performing different analyses. The local machine learning model may contain a model that identifies objects, a model that detects paths, a model to predict the path, and a model that predicts flow. Some examples show that the local machine learning model (216) includes a single, complex model that generates multiple types traffic data.

“In some cases, the local machine-learning model 216 could accept sensor data from cameras and/or sensors as inputs. Local machine learning model 216 can use any number of models including decision trees, logistic regression models and neural networks. It may also use classifiers, support vector models, inductive logic programming, and ensembles of models (e.g. using techniques like bagging, boosting or random forests). Bayesian networks, genetic algorithms, Bayesian models, etc. and can be trained using many approaches such as deep learning. Perceptrons, association rules. Inductive logic. Clustering. Maximum entropy classification. Learning classification. The local machine learning model (216) may use supervised training in some cases. The local machine learning model (216) may use unsupervised learning in some cases.

“The sensor processing module 218 receives data from traffic light 230 and other lights and processes it. There are many sensors that can be used, including cameras that capture images of the intersection. You can use range imaging sensors and microwave sensors as well as inductive loops, IR cameras, IR cameras and multi-spectral cameras. Also, directional sensors, radar, lidar, microphones, audio sensors and/or combinations of sensors. These cameras could be video/photographic cameras, or any other optical sensing device that captures images. The cameras could be set up to capture images at a particular intersection. One camera may capture static images and video of the area. Multiple images are also captured at high speeds (e.g. 30 images per second). Cameras are used to record image data and can be controlled using commands from traffic lights 230, 210, 222, or remote planning systems 222. Cameras capture light and create images. The cameras may use visible light to create images in some cases. Some cameras may use light from other parts of spectrum such as UV light or IR light.

There are many different triggers for sensors and cameras. A passive infrared motion sensor (PIR), may be embedded into cameras. This sensor triggers the cameras to capture one to several images when motion is detected. A microwave motion sensor may also be included in the camera. This sensor triggers the cameras to capture one or several images when motion is detected. You may find the cameras to be ‘normally open? The cameras may be ‘normally open? Digital input that triggers capture of one or more images when sensors communicating with the cameras detect motion, or other events. Some implementations allow the cameras to be commanded to capture images when motion is detected by other sensors or another traffic event. The camera may be commanded by the traffic light 230 or the remote control system 210.

“In some cases, sensors, including cameras trigger external or integrated illuminators (e.g. IR,?white?). lights, etc.) To improve image quality in dark areas. A separate or integrated light sensor can be used to determine whether illumination is required. This may lead to an increase in image quality.

The sensors, including cameras can be programmed with any combination time/day schedules or other variables to determine if images should be captured when triggers occur. When not taking images, the cameras can go into low-power mode. The cameras might wake up periodically in this situation to check for any new messages from the controller. If the cameras are located far from the traffic light 230 or local control unit 210, they may be powered with internal, rechargeable batteries. To recharge the battery, the cameras can use a small solar cell. If the cameras are located together with the controller, they can be powered by the 112 power supply.

The system 200 may include one or more sensors, or detectors. Input from these sensors can be processed by the sensor processing unit 218. The sensors 219 could include any combination of a contact sensor or motion sensor, induction loop sensor or an RGB camera or IR camera. The sensors could include a radio frequency identification (RFID), sensor that identifies an article with a pre-assigned RFID label.

“The local machine-learning model 216 transmits traffic data over the network 220, to the remote planning software 222. An example control system 200 contains a network 220. It could be a local area network, a wide-area network (WAN), or any combination thereof. The network 220 links the local control system (210), the traffic light (230), and the remote planning software (222). The network could include 802.11 Wi-Fi? Wireless Ethernet (e.g. using low-power WiFi chipsets), Bluetooth networks, and wired Ethernet networks. The network 220 could be a mesh network built from the devices that are connected to the mesh network in some cases.

“The remote planning model 222 is located remote from the local machine-learning model 216. It may be a remote plan system, as described in FIG. 1. Multiple local control systems may be used to provide traffic data to the remote planning system 222. The remote planning system 222 could be used as a central control system to manage the entire traffic system in a state. The remote planning system 222 can receive traffic data from different types of intersections under various traffic conditions. Remote planning system 222 uses traffic data from local control systems (210) to generate remote control data.

“The remote planning software 222 includes one or several remote machine learning models 224. The remote machine learning models include one model for every type of analysis. Remote machine learning models 224 could include a model to recognize path patterns, congestion detection, and emergency planning and detection. The remote machine learning models include one model for each intersection that falls under the control of remote planning system 222. Remote machine learning models 224 could include one model per state, one model per district, and one model for every intersection.

The traffic light 230 could be a standard traffic signal that is placed at an intersection to regulate traffic flow. The traffic light 230 could be used as a pedestrian or crosswalk light. The traffic light 230 could include turn signals in some cases. The traffic light 230 could include or be combined with other traffic control devices such as railroad crossing lights and high occupancy vehicle (HOV), lane lights, toll booth lighting, etc.

“The RTOS 232 controls the traffic light 230 in real time. The RTOS 232 is able to receive control instructions from both the remote planning system 222 and the local planning module 212. The RTOS 232 has the function of enforcing safety rules. For example, it may prevent unsafe conditions such as simultaneous green lights on perpendicular roads segments. To comply with safety requirements, the RTOS232 can override local planning module 220 and remote planning system 222.

“In some cases, the remote control instructions are received by the local planning module 212, which then assesses them. If necessary, the local planning module 210 modifies the received instructions to address intersection conditions or passes them on to the RTOS 232. The RTOS 232 carries the instruction from local planning module 210 if safe to do so. Some implementations provide instructions from both the remote planning system 222 and the local planning module 212 to the RTOS 232 through the local control system 210. The RTOS 232 then determines which instructions to follow. Using traffic data from the intersection where the traffic light 230 is located, the local control systems 210 can adjust traffic signal control dynamically. The local control system 210 might determine that an errant motorist has run the red light and may not immediately adjust the phase of traffic lights in the perpendicular directions to green.

“FIG. “FIG. A system like the traffic light control system 200 could perform the process 300.

The process 300 consists of stages (A) to (G). The stages (A) through (G) can be found in the illustrated sequence. However, they could also occur in a different sequence. The process 300 may be repeated in some cases. Some examples only show certain stages of the 300-step process being repeated.

“The 300 process begins at stage (A), where a local control system 210 receives sensor data via sensors 310 that are associated with a specific traffic light 230. The sensors 310 may include a camera. The camera located near the traffic light 230 can provide camera data to the local control system. Some examples include a phase of traffic light 230 in the sensor data. The sensor data might indicate, for example, that the traffic light at 230 is red. In some cases, the sensor data may include a phase that includes other control devices at intersection. An intersection might have multiple traffic lights, pedestrian lights, dynamic traffic signs, and so forth. Some examples may show the phases of the control devices in the sensor data. Other sensors may be used in some cases, such as range imaging sensors and microwave sensors, inductive loops.

“In some cases, the sensor processing module 218 processes sensor and camera data and supplies the processed data to local machine learning model 215. The sensor processing module 218 may process sensor data from the sensors 310 and combine multiple sensor inputs. The sensor processing module 218 might use sensor fusion to synchronize sensor data. The sensor processing module 218 might use filtering techniques such as the Kalman filter or extended Kalman filter, Dempster-Shafer theory and Bayesian networks. “To synchronize sensor inputs.”

“In some cases, the sensor processing module 218 uses machine-learning techniques to analyze and synchronize sensor data. The sensor processing module 218 could accept sensor data from cameras and/or sensors as inputs. The sensor processing module 218 can use any number of models including decision trees, logistic regression models and neural networks. It also supports inductive logic programming. Ensembles of models may be used (e.g. using techniques like bagging, boosting or random forests). You can train the sensor processing module 218 using many methods, including deep learning, perceptrons and association rules, inductive logic. Clustering, maximum entropy, learning classification, and others. The sensor processing module 218 might use supervised learning in some cases. The sensor processing module 218 could use unsupervised learning in some cases.

“The 300th stage continues with stage (B), in which the local control systems 210 uses the local machine-learning model 216 to generate traffic data. The local machine-learning model 216 could operate on data and process it in the same way as the local machine models discussed above. 1. Some examples of the local machine-learning model 216 include multiple machine learning models that can operate simultaneously and separately. The local machine learning model may contain models that can identify pedestrians and obstacles.

“As discussed above in relation to FIG. 2. In some cases, the local machine-learning model 216 could accept sensor data from cameras and/or sensors as inputs. Local machine learning model 216 can use any number of models including decision trees, logistic regression models and neural networks. It may also include classifiers, support vector models, inductive logic programming, ensembles (e.g. using techniques like bagging, boosting or random forests). Bayesian networks, genetic algorithms, Bayesian models, etc. and can be trained using many approaches such as deep learning. Perceptrons, association rules. Inductive logic. Clustering. Maximum entropy classification. Learning classification. The local machine learning model (216) may use supervised training in some cases. The local machine learning model (216) may use unsupervised learning in some cases.

“The local machine-learning model 216 can determine how many vehicles pass through an intersection and which road segments the vehicles originate from. It also may determine which road segments the vehicles travel to. The local machine-learning model 216 can determine the turn fractions of each intersection road segment, as well as paths that are represented by origin-destination pairs. The generated traffic data is then transmitted to the remote control system 210a, which includes the vehicle counts, pedestrian counts, turn fractions, and obstructions. “To the remote planning system 222.

“During stage C, the local control system (210a) transmits the generated traffic data through the network 220 to the remote planning software 222. The network 220, which is the wired network to which each traffic signal 230, local control systems 210, and sensor are connected, can be seen in some examples. The network 220 can be described as a wireless network in some cases. The generated traffic data may be provided by the local control system (210a) using the communication module 214. This module is used to connect the network 220 to remote planning system (222). Traffic data can include raw and processed sensor data from cameras or other sensors. Some examples of traffic data include queue data, path data and vehicle counts.

“In some cases, traffic data may include metadata or features that are extracted from sensor data. Some examples use video, image, or other sensor data to collect data and then process it to determine metadata. The data can then be discarded to reduce storage space requirements. To determine metadata, the incoming sensor data can be analysed at the local control module 220 in substantial real-time. The system can be designed so that the sensor data is not sent to remote planning systems 222 or other remote systems and it is not stored by local control module 209.

This metadata can give valuable insights into traffic patterns as well as past and present conditions. The data could include data related to planning, management and the enactment or modification of laws and policies that affect how people move about. Managers of cities may need information to plan how to design and redesign parts of their cities, for real-time feedback and to actuate; and to evaluate the effects of any changes they make. These data can be collected at different scales: a hamlet or city, a state or continent, and so on. Sensor data from cameras and/or sensors can contain valuable data that is useful for any of these purposes.

“The system 200 can analyze and tag the sensor data to extract valuable metadata. This data can be used without the need for long-term storage (e.g., video feed) This can have many benefits, including greater privacy protection and lower storage costs. The system can anonymize metadata before sending it over the network (e.g. network 230). This ensures that only metadata is sent and not personal data. Users of system 200 can avoid lengthy disputes and embarrassment. A city can’t be called upon to report on a car accident with anonymized metadata and no video feed because the video isn’t saved or made available to other systems.

“The system 200 could use computer vision techniques in order to determine metadata that provides information about how many pedestrians cross at which intersections at what times of day, and where pedestrians are crossing other than at intersections. The system 200 may provide metadata that indicates the frequency of close calls. When pedestrians are at risk and any other environmental factors. The metadata can also indicate where pedestrians cross streets, as well as intersections. The metadata may also indicate incidents and frequency, such as pedestrians walking, cars running red lights, drivers using their cell phones while driving, or counting the incidents. The system 200 can provide metadata that indicates how cyclists interact with traffic and how taxis, buses, and cars interact. The system 200 may provide metadata that indicates how a bus lane, bike lane, or parking lane is used. It can also provide vehicle counts such as bicycle counts and car counts. This metadata can give baseline information about traffic infrastructure usage, which can then be compared to later-generated metadata to determine how changes have affected it. The metadata can show whether changes made to traffic infrastructure, such as a bench or widening a street, have had an impact on the usage.

“The system 200 could extract such metadata using existing infrastructure such as cameras in a city. The system 200 may use machine vision techniques to recognize vehicles and people, and the algorithms and training are tailored to each case. These data can be linked to various actuation systems. The metadata can be used for enforcement or emergency response. The system 200 could be used to supplement existing red-light cameras and CCTV systems. It may also use metadata to direct enforcement officers or first responders to the scene. The system 200 might adjust traffic signals’ timing to accommodate the queue length of vehicles that accumulate at an intersection. Machine vision can be used to count vehicles more accurately than existing sensors such as induction loops. Induction loops often fail to register motorcycles as cars due to the smaller shift of electromagnetic field caused by the motorcycle’s smaller metal mass. The system 200 can be used to account for unusual circumstances, such as prioritizing trucks or buses in HOV lanes, or accommodating pedestrians who are waiting long to cross at busy intersections.

“In some cases, the system 200 could be connected to actuators or dynamic signage to adjust switchable or dynamic lanes based either on traffic along the passageway or predicted traffic due to observed downstream traffic. The system 200 could control electronic traffic signs, such as stop signs that are changed to yield signs when there is less traffic.

“In certain cases, the system 200 might perform real-time routing of public transit to meet demand and traffic. The system 200, for example, may open or close certain lanes in order to direct traffic at high volumes during rush hour. The extracted metadata may be used by the system 200 to input simulation, modeling tools or other planning tools. Vendors and service providers might use the extracted metadata to track consumer reactions to their products. Based on the extracted metadata, vendors and/or service provider may alter what, how much and where they provide products and services. This adjustment may be made in response to immediate circumstances. Some users may use metadata to adjust future planning or make changes based upon historical observations.

“During stage D, the remote planning system 222 receives traffic data generated by local control systems 210a-210d controlling traffic lights at various intersections. Remote planning system 222, which uses traffic data and the remote machine learning model 220, generates a remote instruction for each local control system 210 a-210. This instructs the remote planning system to generate the appropriate remote control instructions.

“In certain implementations, remote planning system 222 may receive traffic data from sensors not located near traffic intersections. The remote planning system 222 could receive traffic data from mobile devices belonging to travellers, for example. The mobile devices of travellers could include a variety of mobile devices such as smart-phones and cellular phones. Traffic data can include the average speed of a vehicle traveling in a specific direction. The traffic data could include, for example, an estimate of the average speed of a vehicle traveling northbound on I-95. Some traffic data can include destination and/or origin data. To supplement the traffic data collected at intersections by sensors including cameras, such as GPS data, or to determine metadata like turn fractions, vehicle speeds, origin-destination pairs and so on, it may be possible to use location data from a mobile device.

“In cases where the systems described here collect personal data about users or make use of that information, users may have the opportunity to choose whether or not programs or features collect this information. This could include information about a user?s location, preferences, and social networks, as well as information about the user’s profession. It may also allow users to control how they receive content from content servers that might be more relevant to them. Certain data can be anonymized before being stored or used. This will ensure that no personally identifiable information is lost. Anonymizing a user’s identity can ensure that no personally identifiable information is available to them. Or, the location information of users may be used to generalize where it is located (such as at a city, ZIP code or state level) so that it cannot be determined where a specific user is located. The user can have control over the information that is collected and used by content servers.

“Some implementations use reinforcement learning to improve and update the remote machine-learning model 224. The remote planning system (222) is shown to receive traffic data generated by multiple local control system 210 a-d. However, remote planning system (222) could also receive traffic data from fewer or more local control system 210.

“As discussed above in relation to FIG. 2. In some cases, remote machine learning model (224) may accept sensor data from cameras and/or sensors as inputs. Remote machine learning model 224, which can use any number of models including decision trees, logistic regression models and neural networks, as well as classifiers, support vector models, inductive logic programming, ensembles (e.g. using techniques like bagging, boosting, random forest, etc.). You can train the model using many methods, including deep learning, perceptrons and association rules. They can also be trained with inductive logic, clustering or maximum entropy classification. The remote machine learning model 224, in some cases, may employ supervised learning. The remote machine learning model, 224 may use unsupervised learning in some cases.

“In some cases, the remote machine-learning model 224 uses reinforcement to decide which actions to take in a specific traffic situation. The remote machine learning model may also use Markov decision processes. The remote machine learning model may employ, for example, the criterion or optimality, brute force or direct policy search or value functions including Monte Carlo or temporal differences methods.

“The remote planning system 222 may update remote machine learning model 228 by calculating new weights to modify inputs (e.g. traffic data) into the neural network of remote machine learning model 228. The remote planning system 222 might determine, based upon the traffic data received from multiple local control systems (210 a-d), that traffic congestion is most severe at that intersection subset. It may also determine that a greater weight should be used to input queue data to the local machine learning program 224.

“A part of updating the remote machine-learning model 224 may include performing experiments or implementing behavior modifications among traffic lights, and assessing the results. The remote machine learning model may, for example, extend the green-light phase by 5 seconds at certain intersections or times. The remote machine learning model (224) can be trained to identify which traffic control patterns work best at specific times and situations. This is possible by introducing different traffic signal operations and monitoring changes in the outcomes. The remote machine learning model can be used to?explore? various techniques. The methods by which data is obtained or used. The remote machine learning model (224) can employ a variety of solutions to the multi-armed bandit issue or the context bandit problem, such as an Epsilon strategy, Epsilon first strategy, Epsilon decreasing strategy or an adaptive strategy. (e.g., greedy algorithms, etc.

“When the remote learning model 228 is updated, it is used to generate remote instructions. The remote instruction can be applied to any intersection within the scope of the remote planning software 222. The remote instruction can be tailored to an intersection controlled via the local control systems 210 or 210.

“The process 300 continues with Stage (E), where the remote planning system 221 transmits the generated remote instructions to the local control systems 210 a via the network 222. The remote planning system 222 could also transmit the generated remote instructions to the local control networks 210 b/d in this example.

“In certain cases, remote instructions can include weight updates to modify the local machine-learning model 216. The remote planning system 222 might determine that turn fractions are not important in controlling traffic flow at intersections controlled locally by local control systems (210 a-d). The remote planning system 222 could transmit a remote instruction to the local machine-learning model 216, including lower weights for inputs regarding turn fractions. Sensor fusion may be used to determine updated weights for inputs into local machine learning models. The sensor processing module 218 might learn information about the reliability and effectiveness of a particular sensor. Local machine learning model (216) may accept sensor data from cameras and/or sensors as inputs. Local machine learning model 216 can use any number of models including decision trees, logistic regression models and neural networks. It may also use classifiers, support vector models, inductive logic programming, ensembles (e.g. using techniques like bagging, boosting or random forests). Bayesian networks, genetic algorithms, Bayesian models, etc. and can be trained using many approaches such as deep learning. Perceptrons, association rules. Inductive logic. Clustering. Maximum entropy classification. Learning classification. The local machine learning model (216) may use supervised training in some cases. The local machine learning model (216) may use unsupervised learning in some cases.

“In some cases, remote instruction may provide new decision making parameters, e.g. weights, thresholds and/or other parameters for the local machine learning model (216), used by the local planning modules 210 a. The local planning module 210a might use a formula to determine the local control instructions to provide traffic light 230. To reduce pedestrian accidents, the remote instruction could include a heavier weight for pedestrian crossing variables in this example. Remote planning system 222, which may include different parameters than the local control systems 210a-210d, can be used to determine the most appropriate. The remote planning system 222 might determine that no cars are traveling east or south through an intersection for which the local controller system 210a is responsible. Remote planning system 222 can also receive local traffic data from local control system210 c, which may indicate that the northern arm’s left turn fraction is 1. The remote planning system may also receive local traffic data from the local controller system 210c. This may indicate that the left turn fraction for the northern arm is 1.

“In some cases, remote instructions may contain control instructions for traffic light 230. The remote instruction could include instructions to keep traffic light 230 on for four minutes, while there is a line at the intersection.

“The 300 process continues with stage F, in which the local controller system 210a makes a local decision on remote instruction and generates local controls instruction. The remote instruction may be ignored by the local control system, 210a when it generates the local control instruction. If there is low accident rate at an intersection, for example, the remote instruction may be ignored by the local control system.

“The 300th stage is completed with stage (G), where the local control instructions generated by local control system 210 a, and the remote instruction generated from remote planning system 222 are transmitted directly to the RTOS 232 which controls traffic light 230. To control the traffic light 230, the RTOS 232 uses both the remote instruction and the local control instruction to generate a signal for traffic control. If the generated traffic data indicates a situation that is not safe, the RTOS 232 can override the local instruction. If a police officer chases a suspect through a red stop sign at an intersection in a direction that is perpendicular with the direction for traffic light 230, the RTOS 232 could ignore remote and local instructions to change from red-to-green.

“Another example is that global traffic data might suggest that because there is heavy traffic heading north on a road, the green light phases at intersections in the northern direction should be extended. However, the RTOS 232 could determine that an emergency vehicle approaches the intersection from the east using local traffic data. The combined remote and local traffic data may allow the RTOS 232 to control the traffic light at the intersection for traffic going north. It will change phase from yellow to red to permit the emergency vehicle to cross.

The control operations described can be performed substantially in real time. The local control modules 210a-210d can provide traffic data at regular intervals. For example, they may deliver information at every 10 seconds, every 30 second, or at another interval. Remote planning system 222 may provide instructions for remote control, although they may not be provided at a set time. The remote planning system 222 can provide instructions for performing a specific pattern of light phase shifts over a period of time or until further instructions are given. The remote planning system 222 can also send instructions dynamically and/or simultaneously to any or all local control units 210a-210d in an area to deal with temporary or unusual conditions.

“In some cases, machine learning models (local and remote) may be substantially updated in real-time. The machine learning models may be updated periodically, for example, every 10 seconds, every 30 second, or at another time. Machine learning models can be updated when control instructions or data are received. This may or not be at a fixed time. Alternately, the machine-learning models can be updated dynamically or asynchronously in response to unusual or temporary conditions.

“In certain cases, local machine learning models might use federated learn. The system 200, for example, updates local machine learning models using weights that are determined by remote planning systems that receive data from multiple intersections. This provides system-level reinforcement learning via the network. To improve traffic flow, generalized rules may be applied to local control system by using the high-level view on traffic conditions.

“Additionally, different intersections may have different local control systems. This allows them to update their local machine learning models to target the most important variables at each intersection. Because each intersection can be controlled by distributed control, traffic flow control may be tailored to suit traffic conditions.

“The remote planning system can test traffic control techniques at multiple intersections. The remote planning system might reduce red light phases at half the intersections within a city. The remote planning system analyzes the changes and makes any necessary adjustments across multiple intersections. If reducing travel time by averaging fifteen minutes for all red light phases, the remote planning software may reverse the red light phase lengths to their original lengths. Remote planning systems may also update remote machine learning models to reflect results from testing. The remote planning system may automate the updating, testing, and analysis of remote machine learning models. The remote planning system might be given a list of variables to modify. The remote planning system might automatically identify variables that can be tested and then perform the tests.

It is possible to model traffic control at the system level and achieve high-level results which improve traffic conditions or assist authorities such as emergency responders. The system could be used to control traffic light phases at the remote planning level to coordinate traffic to clear intersections for a pileup, fire trucks racing to the scene, and police officers following a suspect in high-speed chase, among other things.

“FIG. “FIG. A system like the traffic light control system 200 could perform the process 400.

“A processing device that controls a traffic signal at an intersection receives sensor data (402) which provides images of the intersection. Some examples include data from cameras and/or sensors. Referring to FIG. 2 may receive camera data from cameras located near the traffic signal (230).

“The one or more processing units use one or several local machine learning models to identify the objects at the intersections and paths of objects based upon the sensor data (404). The local control system 210 might use the local machine-learning model 216 to identify vehicles, pedestrians and bicycles at intersections and their respective paths. Based on camera data from cameras located near the traffic light 230.

“The one or more processing units may transmit traffic data from local machine learning models to remote traffic planning systems over a network (406). The communication module 214 may be used by the local control system (210) to transmit the traffic data to remote planning system (222) over the network 220.

“The remote traffic planning system may send a remote instruction to one or more processing devices for the traffic signal that was determined by one or more remote machine-learning models (408). Remote planning system 222 could use remote machine learning model (224) to generate remote instructions. Remote planning system 222 could transmit the remote instruction via the network 220 to local control system210.

“The one or several processing devices may issue a control instruction to the intersection’s traffic signal. This instruction is determined using the remote instruction from remote traffic planning systems and the local instruction generated by one or more processing device (410). The local planning module 212 could generate a local control instructions based on traffic data received from cameras located near the intersection. The remote instruction, local instruction and generated traffic data may be provided by the local control system 210 to the RTOS 232. The RTOS 232 can analyze the remote instruction and local instruction to generate a traffic control instruction. The RTOS 232 could then control traffic light 230 by using the traffic control instruction. The local machine learning model (216) can generate control instructions that are sent directly to the traffic lights 230.

“Data processing apparatus” is a broad term. “Data processing apparatus” can refer to any type of apparatus, devices, or machines that process data. It could include a programmable processor, computer, multiple processors, or computers. An apparatus may include special purpose logic circuitry such as an FPGA (field-programmable gate array), or an ASIC. In addition to hardware, the apparatus may also contain code that creates an environment for the computer program, such as code that is processor firmware, protocol stack, an operating system or a combination thereof.

Summary for “Dynamic Traffic Control”

“This specification generally refers to traffic indicators. Traffic indicators are used to alert vehicles, other objects and travellers along a road. Traffic indicators can be traffic signs, traffic lights, or audio cues to cross. Local controllers control traffic lights by instructing them to change their phase (e.g. from red to green). Traffic congestion is affected by the timing of traffic lights at adjacent intersections. It also affects drivers’ travel time and wait times.

A traffic light control system can regulate traffic lights’ phases at intersections or multiple intersections to maintain smooth and safe traffic flow. Traffic light control systems can identify the appropriate phases of traffic lights using local traffic data at the intersections where they are being controlled and global instructions generated by analyzing traffic data at many other intersections.

“For example, traffic light phases can be coordinated to cut travel time by scheduling traffic light phases so that cars passing through intersections come across a series of green lights or a dynamic green wave. To protect pedestrians crossing intersections, the phases of traffic lights can also be controlled by moving the green light phase at intersections that direct traffic in the pedestrian’s directions. To protect pedestrians, bicyclists and other nonvehicular travellers, it is possible to control lights that inform pedestrians and bicyclists when crossing the intersection is safe. The lights can be programmed to give non-vehicular travelers a lead time (e.g. a headstart) or balance the timing with vehicles who would be passing the non-vehicular traveller.

The traffic light control system controls the phases of traffic lights and can reduce unsafe driving habits. Drivers can be encouraged to follow posted speed limits by having traffic lights at nearby intersections coordinate to change phase. Coordinated traffic lights reduce congestion and reduce the need for roads to be expanded to handle disproportionately high volumes of traffic during peak travel times. They also reduce fuel consumption and vehicle pollution. The traffic light control system reduces frustration among drivers by reducing travel times.

“The combination global instructions and local data allow the traffic control system learn from traffic situations at multiple intersections, while still being flexible based on local conditions. Global instructions might suggest that traffic can be reduced by reducing red light phases at traffic lights. Global instructions can be helpful for many intersections to alleviate congestion. Local data may indicate that a bicyclist crosses an intersection. If the traffic light changes to a green phase, the bicyclist won’t be able finish crossing the intersection. Traffic control systems may alter the instructions to the traffic lights to prolong the red light phase until the bicyclist crosses the intersection. Separate lights such as the walk lights for pedestrians are sometimes used to protect non-vehicular travelers.

“Accordingly to an innovative aspect of this specification, a method of controlling traffic lights is provided. In this system, sensor data providing images of traffic intersections is collected by cameras or other sensors located near the intersection. The sensor data may be collected by sensors at an intersection and used to identify objects using local machine learning models. The processing devices might determine the object’s position relative to the intersection, speed, and direction of travel. Traffic planning systems located far from intersections may be able to receive traffic data including identified objects from local machine learning models. This information can be transmitted over a network. The instructions may be generated by the remote planning system and sent to the processing devices controlling the traffic light. The remote planning system may send instructions to the processing devices controlling the intersection’s traffic lights. These instructions can be used to generate local control instructions. The remote planning system instructions and local instructions may be used by a traffic light controller to generate and give a control instruction.

One innovative aspect of this specification is that it includes receiving camera data from intersections, which can then be used to control traffic signals at intersections. The one or two processing devices are located near the intersection. They also use local machine-learning models to identify objects at intersections and their paths. A remote instruction for traffic signal control at intersections is sent to remote traffic planning systems over a network. This remote instruction is determined based (i) on remote instructions from remote traffic planning systems and (ii) on local instructions generated by one or several processing device.

Implementations can include any or all of the following features. The method could include, for example, accessing safety rules that indicate requirements for traffic signals, modifiying the remote instruction based upon the safety rules, and then providing the modified instruction to traffic signal by one or more processing devices. The method may include determining a condition at an intersection using one or more of the local machine learning models, altering the remote instruction based upon the local instruction and providing the control instruction to traffic signal.

“In some cases, the method involves determining by one or more processors and using one or two local machine learning models, and further based upon the camera data and phase information for the traffic signals indicating a traffic signal’s current phase, a classification of a condition at intersection, and generating by one or more processors a local instruction based the classification and overriding, depending on the remote instruction. The method may include determining the number of objects traversing a path through an intersection using one or more processing units. This includes counting the objects of a specific type that traverse the path. Claim 5, wherein the method includes providing traffic data that indicates the speed at which the counted objects are moving.

“In some cases, providing traffic data may include providing one or more measures that indicate relative numbers of vehicles using different routes through the intersection. Some examples show reinforcement learning, where traffic data from nearby intersections is received and taken into consideration by the remote planning system.

“In some cases, dynamically altering the traffic signal by one or more processing device and based upon the traffic instruction is part of the method. The one or more processing units are usually located at the intersection in some cases. The method may include determining the location of the one or more processing devices based at least partially on traffic data. The method may include providing traffic data from local machine learning models over a network to a remote traffic planning software via the one or several processing devices periodically.

“In some cases, the method may include receiving custom machine learning model parameters for a local machine learning program. The method may include determining the control instructions to the traffic signal at an intersection using a real-time operational system. This is based on safety rules. The method may include using one or more local machine-learning models to classify an intersection condition or assess the speed of an identified object. The method may include storing parameters to generate local control instructions by one or several processors in a database. Accessing the database to retrieve these parameters by one or two processors allows them access the parameters. Finally, they can generate a local control instruction based on the outputs of one or multiple local machine learning models.

“The subject matter in this specification may be implemented in specific embodiments to achieve one or more of these advantages. This system is flexible and can be used with sensor systems such a lidar radar, range imaging microwave sensors, inductive loops black and white cameras, color cameras like red green (RB) cameras, infrared cameras, hyper-spectral camera, multi-spectral cameras (IR) cameras, directional sensors. Other types of data can be collected in some implementations. Data associated with traffic indicators may include data about railroad crossings and traffic signals, such as traffic lights, railroad crossings, flags, signs, road barriers, flags and traffic signs. Data may also be collected. You may also use other sensors or combinations of sensors. This system decreases the wait time for vehicles at intersections by optimizing and controlling traffic lights to control traffic flow. Safety and traffic throughput are also improved. It can also reduce the number and frequency of stops that any traveller (vehicle or otherwise) must make at traffic lights. The system can analyze traffic data to better understand traffic flows for specific modes by analysing them. The system can help you understand the traffic patterns of specific modes. For instance, pedestrians might have a different flow pattern to buses than they do to passengers cars. Monitoring traffic behavior can help to improve traffic safety at intersections. The system can identify factors that could lead to accidents or near-collisions. Signal timing can be used to give priority to certain travelers in some cases. The system can allow pedestrians to cross the road before bicyclists or other non-vehicular travellers. Additional lights such as a warning lamp can be used in certain cases. A warning light can be added to a traffic signal, for example, to warn drivers when a collision is imminent.

“The accompanying drawings and description below detail one or more implementations for the subject matter. The claims, drawings, and description will reveal other potential features, aspects, or advantages to the subject matter.

“Some traffic control systems include multiple levels of machine-learning processing. A remote system (e.g., server system or cloud-based) may use reinforcement learning for traffic control coordination among multiple traffic lights within an area, such as a city. Local processing systems that run other machine learning models may also be distributed in the area. For example, a local system can be installed at each intersection. The local systems monitor intersection conditions and send traffic data to remote systems. Remote systems use its learning models to generate instructions which are then sent to local systems. This arrangement allows traffic lights to be optimized or coordinated at the network level. However, it also gives local intersections the ability to adapt to local conditions using local machine learning models. The remote system’s instructions may improve traffic throughput and efficiency over large areas. However, each local processing unit might be able override these instructions to address local conditions, such as elderly people crossing the street or emergency vehicles approaching.

“In some cases, a traffic control systems controls a traffic light at an intersection. A local processing system runs a real time operating system (RTOS) and controls each traffic light. The local processing system may be located near the intersection where the traffic lights are located. Other modules within the local processing network may send instructions to the RTOS, including a local planning module, which is part of the local system, or a remote planning system, which sends data over the internet. Local machine learning models may be used by the local planning module to identify intersection objects and generate traffic control instructions based on the local conditions. Remote machine learning models may be used by the remote planning system to generate traffic light control instructions based on conditions at different intersections. The remote planning system could receive traffic data from multiple intersections and use reinforcement learning to improve traffic flow over a large area with many intersections.

“FIG. “FIG. A traffic intersection is generally composed of one or more traffic signals and connecting two or more roads. The techniques described in this article can be used to control any traffic signal. A traffic intersection is a three-way intersection, which connects one end of a road with another length. It can have three traffic lights, one for each segment of road or arm. It can also have arms extending from it. Traffic intersections may have more or less arms depending on how they are implemented. This particular example shows a four-way intersection 100 that connects two roads. Each arm of the intersection is represented by a separate traffic light, 102 a-d.

In some cases, different sets of lights may be included at each intersection’s arm. An intersection might have a number of lights, such as one for bicyclists and another for passengers. Traffic lights may include audio cues as an alternative or addition to visual cues. Traffic lights might include a pedestrian light and speaker output to indicate that it is safe to cross. Traffic lights may include accessibility features such as vibration modules or other similar technologies.

“In this example, several sensors, including camera104 a?d and other sensors106 a?d, are shown as being proximate each traffic light (102 a?d). Cameras 104 capture information such as images or other sensor data for each arm. To capture information about an intersection, multiple cameras or sensors can be placed in different locations, such as at traffic lights that are not immediately adjacent to them. There may be several sensors per traffic light in some cases. A single camera or sensor may capture information about multiple lanes, arms or areas of a roadway. Multiple sensors are unnecessary. There may be multiple cameras for each traffic light in some cases.

“The traffic data captured at each traffic light 102 a-d can be transmitted to a local control network 110. In some cases, each intersection is controlled by a local control system. One or more intersections may be controlled by a local control system. A local control system could control a group of intersections that are close by. A local control system could control multiple intersections that are located far from one another. A local control system may be responsible for controlling a specific section of an intersection. A local control system might control one of the intersection’s four traffic lights, while another local control system could control the intersection’s other three traffic lights. Different local control systems could control different numbers.

“In some cases, the local control systems 110 may include one or more local machine-learning models 112. To process traffic data, the local machine learning models 112 will be used. The local machine learning models 112 may store data in a local control system’s memory. The local control system 110 may store metadata that is determined by local machine learning models 112 rather than sensor data from cameras 104 or sensors 106. The local control system 110 might store, for example, the number and speed of vehicles turning right at the intersection’s western arm, as well as the number and speed of vehicles passing through the intersection within the last five minutes.

“The local machine-learning models 112 could accept inputs from cameras and/or sensors. Local machine learning models 112 can use any number of models including decision trees, logistic regression models and neural networks. They also may use classifiers, support vector models, inductive logic programming, ensembles (e.g. using techniques like bagging, boosting or random forests). Bayesian networks, genetic algorithms, Bayesian network, etc. and can be trained using many approaches such as deep learning. Perceptrons, association rules. Inductive logic. Clustering. Maximum entropy classification. Learning classification. Some local machine learning models 112 might use supervised learning. Some local machine learning models 112 may use unsupervised learning in some cases.

“Some examples show how local machine learning models 112 can be used to identify objects at an intersection based upon traffic data. The local machine learning models 112 can also be trained to recognize specific types of objects (e.g., vehicles, busses, trucks, etc.). ), bicycles, pedestrians, debris, etc. The local machine learning model 112 can identify specific details such as make, model, license plate number and other information. characteristics of a vehicle, such as the make, model, license plate number and pedestrian characteristics. It can also identify the number of horses pulling a carriage. It can also determine the size of any debris or obstructions. The local machine learning models 112 can also be trained to recognize traffic scenes such as accidents, construction, pedestrians waiting to cross, emergency vehicles approaching, and so on. The local machine learning models 112 are able to recognize scenes and objects in certain situations. Both within the intersection and adjacent streets can be used to train the local machine learning models 112 to recognize objects, scenes, etc. The local machine learning models 112 can also be used to detect an emergency vehicle coming from an adjacent street. The local machine learning model 112 can be used to determine the speed, direction and acceleration of vehicles that pass through the intersection.

“The local machine-learning models 112 can determine the number of vehicles that pass through an intersection from a particular path. The local machine learning models 112 can identify that five vehicles travel from the western segment of intersection and that three are traveling from the southern segment. No vehicles are coming from either the eastern or northern segments of intersection. Local machine learning models 112 can calculate the speed, direction, acceleration, and acceleration of vehicles that pass through the intersection.

“The local machine-learning models 112 are able to identify traffic patterns such as the formation of queues. The local machine learning model 112 can, for example, determine that there is a lot of traffic because there are many vehicles waiting to get in the intersection. The local machine learning models 112 could also use camera data to determine if there is a large queue of vehicles waiting for exit. Sometimes, an intersection might be blocked so that no traffic is moving through it, even though the traffic lights are green. This could happen when traffic is blocked ahead of the intersection so that cars cannot pass through it. These queue data can be stored and generated to help determine if the queue conditions have cleared. Some examples show that queue data can be used to input the local learning models 112 in order to update control instructions or improve pattern recognition.

“The local machine-learning models 112 can identify the paths taken by identified vehicles. The identified paths can be described as origin-destination pair, which indicates the origin and destination of each vehicle. Some examples show the origin-destination pair as vectors. You can organize the origin-destination pairs in many ways, including lists and tables.

“The local control systems 110 generate local control instructions using traffic data collected by cameras at intersections. Local control system 110 generates local control instructions using local machine learning models 112. The traffic data may be used to customize the local control instructions for each traffic light. The local control system might receive traffic data that indicates that very few vehicles are traveling from the intersection’s northern or southern roads segments, but many are using the eastern and western road segments. The local control system 110 might generate an instruction to increase the green phase at the western and eastern segments of an intersection. It may also generate an instruction to shorten the green phase at the northern or southern segments of an intersection. The local control instructions may be generalized to the intersections or portions of intersections for which the local control system is responsible. The local control system 110 might generate instructions to reduce the yellow-to red phase transition time.

“The traffic data generated from the local machine-learning models 112 are transmitted to a remote traffic management system 122. The remote planning system 122 is connected to the local control system 110 via a network 120.

Remote planning system 122 may include separate remote machine-learning models 124. Remote planning system 122 can receive data from multiple intersections. Remote machine learning models (124) use traffic data from multiple intersections to determine global traffic control directions. Remote machine learning models 124 might use traffic data from multiple intersections to determine that the yellow-to red light transition at each intersection within the district needs to be increased by two seconds.

Remote machine learning models 124 can accept sensor data from cameras and/or sensors. Local machine learning models 112 can use any number of models including decision trees, logistic regression models and neural networks. They also may use classifiers, support vector models, inductive logic programming, ensembles (e.g. using techniques like bagging, boosting or random forests). You can train them using many methods, including deep learning, perceptrons and association rules. They can also be trained with inductive logic, clustering or maximum entropy classification. The remote machine learning models may employ supervised learning in some cases. Remote machine learning models 124 may use unsupervised learning in some cases.

“In some cases, remote machine learning models (124) uses reinforcement learning to decide which actions to take in particular traffic situations. The remote machine learning models may also use Markov decision processes. Remote machine learning models 124 could use, for example, the criterion or optimality, brute force or direct policy search or value functions including Monte Carlo or temporal differences methods.

“The remote planning software 122 can use the global traffic control instruction to generate control instructions that are transmitted to the local control system 110. The control instructions may be identical to the global traffic control instruction in some cases. The remote planning system 122 control instructions can be customized to suit each traffic light that the local control system 110 controls.

“The remote control system 122 transmits remote control instructions through the network 120 to the local control systems 110. The remote control system 122 may be used to generate the local control instructions 110. Sometimes, local control instructions can override or modify remote control instructions. If the remote control instructions reduce the length of the green-light phase but the intersection the local control 110 is monitoring is experiencing delays, the local system could instead generate local control instruction that increase the green-light phase at the intersection.

“The local control system contains an RTOS 130 which controls the traffic lights 102 a?d. Based on the remote control instruction and the local control instruction and traffic data 100 from the cameras 104 a?d and sensors 106 a?d, the RTOS 130 generates a traffic light control instruction for each traffic light 102 a?d.

“FIG. “FIG. The control system 200 comprises a local control network 210, a network 220 and remote planning system 223. A traffic light 230 is also included.

“The local control systems 210 could be the same as those described in FIG. 1. Local control system 210 contains a local planning module 212 and a communication module 214. A local machine learning model 216 is also included. The sensor processing module 218 and an RTOS 232 are also part of the local control system 210.

“The local planning module (212) receives traffic data via the local machine-learning model 216. The local planning module 212 receives traffic data and generates control instructions for traffic lights. The local planning module 212 might receive traffic data that indicates that no vehicles are present at the intersection’s western, eastern, and southern roads segments. The local planning module 212 may extend the green phase of the traffic light for northern road segments of the intersection.

“The communication module 214, which is also known as the communication module 220, allows for communication over a network (e.g. the network 220-2) and with traffic lights (e.g. the traffic light 233. The communication module 214, which can be wirelessly connected, allows for wireless communication over the network 220. The communication module 214 could be used to transmit data over wireless channels and wireless voice channels. The communication module 214 can transmit traffic data via a wireless data channel. One or more LTE modules, a GSM modul, a radio module, a cellular transmission module or any other type of module that can exchange communications in one the following formats: LTE or GPRS or CDMA, EDGE, EDGE, EVDO, UMTS or IP.

“The communication module (214) may also be a wired module that allows for communication over the network 220 via a wired connection. The communication module could be, for example, a modem or a network interface card. A communication module could be an Ethernet network card that allows the monitoring control unit 110 and the monitoring system 110 to communicate over local area networks or the Internet. The communication module 214 can be set up to communicate with traffic lights specific protocols.

“The local machine-learning model 216 receives data directly from cameras located at intersections or intersections where the local control system210 is responsible. The local machine learning model (216) may include multiple machine learning models performing different analyses. The local machine learning model may contain a model that identifies objects, a model that detects paths, a model to predict the path, and a model that predicts flow. Some examples show that the local machine learning model (216) includes a single, complex model that generates multiple types traffic data.

“In some cases, the local machine-learning model 216 could accept sensor data from cameras and/or sensors as inputs. Local machine learning model 216 can use any number of models including decision trees, logistic regression models and neural networks. It may also use classifiers, support vector models, inductive logic programming, and ensembles of models (e.g. using techniques like bagging, boosting or random forests). Bayesian networks, genetic algorithms, Bayesian models, etc. and can be trained using many approaches such as deep learning. Perceptrons, association rules. Inductive logic. Clustering. Maximum entropy classification. Learning classification. The local machine learning model (216) may use supervised training in some cases. The local machine learning model (216) may use unsupervised learning in some cases.

“The sensor processing module 218 receives data from traffic light 230 and other lights and processes it. There are many sensors that can be used, including cameras that capture images of the intersection. You can use range imaging sensors and microwave sensors as well as inductive loops, IR cameras, IR cameras and multi-spectral cameras. Also, directional sensors, radar, lidar, microphones, audio sensors and/or combinations of sensors. These cameras could be video/photographic cameras, or any other optical sensing device that captures images. The cameras could be set up to capture images at a particular intersection. One camera may capture static images and video of the area. Multiple images are also captured at high speeds (e.g. 30 images per second). Cameras are used to record image data and can be controlled using commands from traffic lights 230, 210, 222, or remote planning systems 222. Cameras capture light and create images. The cameras may use visible light to create images in some cases. Some cameras may use light from other parts of spectrum such as UV light or IR light.

There are many different triggers for sensors and cameras. A passive infrared motion sensor (PIR), may be embedded into cameras. This sensor triggers the cameras to capture one to several images when motion is detected. A microwave motion sensor may also be included in the camera. This sensor triggers the cameras to capture one or several images when motion is detected. You may find the cameras to be ‘normally open? The cameras may be ‘normally open? Digital input that triggers capture of one or more images when sensors communicating with the cameras detect motion, or other events. Some implementations allow the cameras to be commanded to capture images when motion is detected by other sensors or another traffic event. The camera may be commanded by the traffic light 230 or the remote control system 210.

“In some cases, sensors, including cameras trigger external or integrated illuminators (e.g. IR,?white?). lights, etc.) To improve image quality in dark areas. A separate or integrated light sensor can be used to determine whether illumination is required. This may lead to an increase in image quality.

The sensors, including cameras can be programmed with any combination time/day schedules or other variables to determine if images should be captured when triggers occur. When not taking images, the cameras can go into low-power mode. The cameras might wake up periodically in this situation to check for any new messages from the controller. If the cameras are located far from the traffic light 230 or local control unit 210, they may be powered with internal, rechargeable batteries. To recharge the battery, the cameras can use a small solar cell. If the cameras are located together with the controller, they can be powered by the 112 power supply.

The system 200 may include one or more sensors, or detectors. Input from these sensors can be processed by the sensor processing unit 218. The sensors 219 could include any combination of a contact sensor or motion sensor, induction loop sensor or an RGB camera or IR camera. The sensors could include a radio frequency identification (RFID), sensor that identifies an article with a pre-assigned RFID label.

“The local machine-learning model 216 transmits traffic data over the network 220, to the remote planning software 222. An example control system 200 contains a network 220. It could be a local area network, a wide-area network (WAN), or any combination thereof. The network 220 links the local control system (210), the traffic light (230), and the remote planning software (222). The network could include 802.11 Wi-Fi? Wireless Ethernet (e.g. using low-power WiFi chipsets), Bluetooth networks, and wired Ethernet networks. The network 220 could be a mesh network built from the devices that are connected to the mesh network in some cases.

“The remote planning model 222 is located remote from the local machine-learning model 216. It may be a remote plan system, as described in FIG. 1. Multiple local control systems may be used to provide traffic data to the remote planning system 222. The remote planning system 222 could be used as a central control system to manage the entire traffic system in a state. The remote planning system 222 can receive traffic data from different types of intersections under various traffic conditions. Remote planning system 222 uses traffic data from local control systems (210) to generate remote control data.

“The remote planning software 222 includes one or several remote machine learning models 224. The remote machine learning models include one model for every type of analysis. Remote machine learning models 224 could include a model to recognize path patterns, congestion detection, and emergency planning and detection. The remote machine learning models include one model for each intersection that falls under the control of remote planning system 222. Remote machine learning models 224 could include one model per state, one model per district, and one model for every intersection.

The traffic light 230 could be a standard traffic signal that is placed at an intersection to regulate traffic flow. The traffic light 230 could be used as a pedestrian or crosswalk light. The traffic light 230 could include turn signals in some cases. The traffic light 230 could include or be combined with other traffic control devices such as railroad crossing lights and high occupancy vehicle (HOV), lane lights, toll booth lighting, etc.

“The RTOS 232 controls the traffic light 230 in real time. The RTOS 232 is able to receive control instructions from both the remote planning system 222 and the local planning module 212. The RTOS 232 has the function of enforcing safety rules. For example, it may prevent unsafe conditions such as simultaneous green lights on perpendicular roads segments. To comply with safety requirements, the RTOS232 can override local planning module 220 and remote planning system 222.

“In some cases, the remote control instructions are received by the local planning module 212, which then assesses them. If necessary, the local planning module 210 modifies the received instructions to address intersection conditions or passes them on to the RTOS 232. The RTOS 232 carries the instruction from local planning module 210 if safe to do so. Some implementations provide instructions from both the remote planning system 222 and the local planning module 212 to the RTOS 232 through the local control system 210. The RTOS 232 then determines which instructions to follow. Using traffic data from the intersection where the traffic light 230 is located, the local control systems 210 can adjust traffic signal control dynamically. The local control system 210 might determine that an errant motorist has run the red light and may not immediately adjust the phase of traffic lights in the perpendicular directions to green.

“FIG. “FIG. A system like the traffic light control system 200 could perform the process 300.

The process 300 consists of stages (A) to (G). The stages (A) through (G) can be found in the illustrated sequence. However, they could also occur in a different sequence. The process 300 may be repeated in some cases. Some examples only show certain stages of the 300-step process being repeated.

“The 300 process begins at stage (A), where a local control system 210 receives sensor data via sensors 310 that are associated with a specific traffic light 230. The sensors 310 may include a camera. The camera located near the traffic light 230 can provide camera data to the local control system. Some examples include a phase of traffic light 230 in the sensor data. The sensor data might indicate, for example, that the traffic light at 230 is red. In some cases, the sensor data may include a phase that includes other control devices at intersection. An intersection might have multiple traffic lights, pedestrian lights, dynamic traffic signs, and so forth. Some examples may show the phases of the control devices in the sensor data. Other sensors may be used in some cases, such as range imaging sensors and microwave sensors, inductive loops.

“In some cases, the sensor processing module 218 processes sensor and camera data and supplies the processed data to local machine learning model 215. The sensor processing module 218 may process sensor data from the sensors 310 and combine multiple sensor inputs. The sensor processing module 218 might use sensor fusion to synchronize sensor data. The sensor processing module 218 might use filtering techniques such as the Kalman filter or extended Kalman filter, Dempster-Shafer theory and Bayesian networks. “To synchronize sensor inputs.”

“In some cases, the sensor processing module 218 uses machine-learning techniques to analyze and synchronize sensor data. The sensor processing module 218 could accept sensor data from cameras and/or sensors as inputs. The sensor processing module 218 can use any number of models including decision trees, logistic regression models and neural networks. It also supports inductive logic programming. Ensembles of models may be used (e.g. using techniques like bagging, boosting or random forests). You can train the sensor processing module 218 using many methods, including deep learning, perceptrons and association rules, inductive logic. Clustering, maximum entropy, learning classification, and others. The sensor processing module 218 might use supervised learning in some cases. The sensor processing module 218 could use unsupervised learning in some cases.

“The 300th stage continues with stage (B), in which the local control systems 210 uses the local machine-learning model 216 to generate traffic data. The local machine-learning model 216 could operate on data and process it in the same way as the local machine models discussed above. 1. Some examples of the local machine-learning model 216 include multiple machine learning models that can operate simultaneously and separately. The local machine learning model may contain models that can identify pedestrians and obstacles.

“As discussed above in relation to FIG. 2. In some cases, the local machine-learning model 216 could accept sensor data from cameras and/or sensors as inputs. Local machine learning model 216 can use any number of models including decision trees, logistic regression models and neural networks. It may also include classifiers, support vector models, inductive logic programming, ensembles (e.g. using techniques like bagging, boosting or random forests). Bayesian networks, genetic algorithms, Bayesian models, etc. and can be trained using many approaches such as deep learning. Perceptrons, association rules. Inductive logic. Clustering. Maximum entropy classification. Learning classification. The local machine learning model (216) may use supervised training in some cases. The local machine learning model (216) may use unsupervised learning in some cases.

“The local machine-learning model 216 can determine how many vehicles pass through an intersection and which road segments the vehicles originate from. It also may determine which road segments the vehicles travel to. The local machine-learning model 216 can determine the turn fractions of each intersection road segment, as well as paths that are represented by origin-destination pairs. The generated traffic data is then transmitted to the remote control system 210a, which includes the vehicle counts, pedestrian counts, turn fractions, and obstructions. “To the remote planning system 222.

“During stage C, the local control system (210a) transmits the generated traffic data through the network 220 to the remote planning software 222. The network 220, which is the wired network to which each traffic signal 230, local control systems 210, and sensor are connected, can be seen in some examples. The network 220 can be described as a wireless network in some cases. The generated traffic data may be provided by the local control system (210a) using the communication module 214. This module is used to connect the network 220 to remote planning system (222). Traffic data can include raw and processed sensor data from cameras or other sensors. Some examples of traffic data include queue data, path data and vehicle counts.

“In some cases, traffic data may include metadata or features that are extracted from sensor data. Some examples use video, image, or other sensor data to collect data and then process it to determine metadata. The data can then be discarded to reduce storage space requirements. To determine metadata, the incoming sensor data can be analysed at the local control module 220 in substantial real-time. The system can be designed so that the sensor data is not sent to remote planning systems 222 or other remote systems and it is not stored by local control module 209.

This metadata can give valuable insights into traffic patterns as well as past and present conditions. The data could include data related to planning, management and the enactment or modification of laws and policies that affect how people move about. Managers of cities may need information to plan how to design and redesign parts of their cities, for real-time feedback and to actuate; and to evaluate the effects of any changes they make. These data can be collected at different scales: a hamlet or city, a state or continent, and so on. Sensor data from cameras and/or sensors can contain valuable data that is useful for any of these purposes.

“The system 200 can analyze and tag the sensor data to extract valuable metadata. This data can be used without the need for long-term storage (e.g., video feed) This can have many benefits, including greater privacy protection and lower storage costs. The system can anonymize metadata before sending it over the network (e.g. network 230). This ensures that only metadata is sent and not personal data. Users of system 200 can avoid lengthy disputes and embarrassment. A city can’t be called upon to report on a car accident with anonymized metadata and no video feed because the video isn’t saved or made available to other systems.

“The system 200 could use computer vision techniques in order to determine metadata that provides information about how many pedestrians cross at which intersections at what times of day, and where pedestrians are crossing other than at intersections. The system 200 may provide metadata that indicates the frequency of close calls. When pedestrians are at risk and any other environmental factors. The metadata can also indicate where pedestrians cross streets, as well as intersections. The metadata may also indicate incidents and frequency, such as pedestrians walking, cars running red lights, drivers using their cell phones while driving, or counting the incidents. The system 200 can provide metadata that indicates how cyclists interact with traffic and how taxis, buses, and cars interact. The system 200 may provide metadata that indicates how a bus lane, bike lane, or parking lane is used. It can also provide vehicle counts such as bicycle counts and car counts. This metadata can give baseline information about traffic infrastructure usage, which can then be compared to later-generated metadata to determine how changes have affected it. The metadata can show whether changes made to traffic infrastructure, such as a bench or widening a street, have had an impact on the usage.

“The system 200 could extract such metadata using existing infrastructure such as cameras in a city. The system 200 may use machine vision techniques to recognize vehicles and people, and the algorithms and training are tailored to each case. These data can be linked to various actuation systems. The metadata can be used for enforcement or emergency response. The system 200 could be used to supplement existing red-light cameras and CCTV systems. It may also use metadata to direct enforcement officers or first responders to the scene. The system 200 might adjust traffic signals’ timing to accommodate the queue length of vehicles that accumulate at an intersection. Machine vision can be used to count vehicles more accurately than existing sensors such as induction loops. Induction loops often fail to register motorcycles as cars due to the smaller shift of electromagnetic field caused by the motorcycle’s smaller metal mass. The system 200 can be used to account for unusual circumstances, such as prioritizing trucks or buses in HOV lanes, or accommodating pedestrians who are waiting long to cross at busy intersections.

“In some cases, the system 200 could be connected to actuators or dynamic signage to adjust switchable or dynamic lanes based either on traffic along the passageway or predicted traffic due to observed downstream traffic. The system 200 could control electronic traffic signs, such as stop signs that are changed to yield signs when there is less traffic.

“In certain cases, the system 200 might perform real-time routing of public transit to meet demand and traffic. The system 200, for example, may open or close certain lanes in order to direct traffic at high volumes during rush hour. The extracted metadata may be used by the system 200 to input simulation, modeling tools or other planning tools. Vendors and service providers might use the extracted metadata to track consumer reactions to their products. Based on the extracted metadata, vendors and/or service provider may alter what, how much and where they provide products and services. This adjustment may be made in response to immediate circumstances. Some users may use metadata to adjust future planning or make changes based upon historical observations.

“During stage D, the remote planning system 222 receives traffic data generated by local control systems 210a-210d controlling traffic lights at various intersections. Remote planning system 222, which uses traffic data and the remote machine learning model 220, generates a remote instruction for each local control system 210 a-210. This instructs the remote planning system to generate the appropriate remote control instructions.

“In certain implementations, remote planning system 222 may receive traffic data from sensors not located near traffic intersections. The remote planning system 222 could receive traffic data from mobile devices belonging to travellers, for example. The mobile devices of travellers could include a variety of mobile devices such as smart-phones and cellular phones. Traffic data can include the average speed of a vehicle traveling in a specific direction. The traffic data could include, for example, an estimate of the average speed of a vehicle traveling northbound on I-95. Some traffic data can include destination and/or origin data. To supplement the traffic data collected at intersections by sensors including cameras, such as GPS data, or to determine metadata like turn fractions, vehicle speeds, origin-destination pairs and so on, it may be possible to use location data from a mobile device.

“In cases where the systems described here collect personal data about users or make use of that information, users may have the opportunity to choose whether or not programs or features collect this information. This could include information about a user?s location, preferences, and social networks, as well as information about the user’s profession. It may also allow users to control how they receive content from content servers that might be more relevant to them. Certain data can be anonymized before being stored or used. This will ensure that no personally identifiable information is lost. Anonymizing a user’s identity can ensure that no personally identifiable information is available to them. Or, the location information of users may be used to generalize where it is located (such as at a city, ZIP code or state level) so that it cannot be determined where a specific user is located. The user can have control over the information that is collected and used by content servers.

“Some implementations use reinforcement learning to improve and update the remote machine-learning model 224. The remote planning system (222) is shown to receive traffic data generated by multiple local control system 210 a-d. However, remote planning system (222) could also receive traffic data from fewer or more local control system 210.

“As discussed above in relation to FIG. 2. In some cases, remote machine learning model (224) may accept sensor data from cameras and/or sensors as inputs. Remote machine learning model 224, which can use any number of models including decision trees, logistic regression models and neural networks, as well as classifiers, support vector models, inductive logic programming, ensembles (e.g. using techniques like bagging, boosting, random forest, etc.). You can train the model using many methods, including deep learning, perceptrons and association rules. They can also be trained with inductive logic, clustering or maximum entropy classification. The remote machine learning model 224, in some cases, may employ supervised learning. The remote machine learning model, 224 may use unsupervised learning in some cases.

“In some cases, the remote machine-learning model 224 uses reinforcement to decide which actions to take in a specific traffic situation. The remote machine learning model may also use Markov decision processes. The remote machine learning model may employ, for example, the criterion or optimality, brute force or direct policy search or value functions including Monte Carlo or temporal differences methods.

“The remote planning system 222 may update remote machine learning model 228 by calculating new weights to modify inputs (e.g. traffic data) into the neural network of remote machine learning model 228. The remote planning system 222 might determine, based upon the traffic data received from multiple local control systems (210 a-d), that traffic congestion is most severe at that intersection subset. It may also determine that a greater weight should be used to input queue data to the local machine learning program 224.

“A part of updating the remote machine-learning model 224 may include performing experiments or implementing behavior modifications among traffic lights, and assessing the results. The remote machine learning model may, for example, extend the green-light phase by 5 seconds at certain intersections or times. The remote machine learning model (224) can be trained to identify which traffic control patterns work best at specific times and situations. This is possible by introducing different traffic signal operations and monitoring changes in the outcomes. The remote machine learning model can be used to?explore? various techniques. The methods by which data is obtained or used. The remote machine learning model (224) can employ a variety of solutions to the multi-armed bandit issue or the context bandit problem, such as an Epsilon strategy, Epsilon first strategy, Epsilon decreasing strategy or an adaptive strategy. (e.g., greedy algorithms, etc.

“When the remote learning model 228 is updated, it is used to generate remote instructions. The remote instruction can be applied to any intersection within the scope of the remote planning software 222. The remote instruction can be tailored to an intersection controlled via the local control systems 210 or 210.

“The process 300 continues with Stage (E), where the remote planning system 221 transmits the generated remote instructions to the local control systems 210 a via the network 222. The remote planning system 222 could also transmit the generated remote instructions to the local control networks 210 b/d in this example.

“In certain cases, remote instructions can include weight updates to modify the local machine-learning model 216. The remote planning system 222 might determine that turn fractions are not important in controlling traffic flow at intersections controlled locally by local control systems (210 a-d). The remote planning system 222 could transmit a remote instruction to the local machine-learning model 216, including lower weights for inputs regarding turn fractions. Sensor fusion may be used to determine updated weights for inputs into local machine learning models. The sensor processing module 218 might learn information about the reliability and effectiveness of a particular sensor. Local machine learning model (216) may accept sensor data from cameras and/or sensors as inputs. Local machine learning model 216 can use any number of models including decision trees, logistic regression models and neural networks. It may also use classifiers, support vector models, inductive logic programming, ensembles (e.g. using techniques like bagging, boosting or random forests). Bayesian networks, genetic algorithms, Bayesian models, etc. and can be trained using many approaches such as deep learning. Perceptrons, association rules. Inductive logic. Clustering. Maximum entropy classification. Learning classification. The local machine learning model (216) may use supervised training in some cases. The local machine learning model (216) may use unsupervised learning in some cases.

“In some cases, remote instruction may provide new decision making parameters, e.g. weights, thresholds and/or other parameters for the local machine learning model (216), used by the local planning modules 210 a. The local planning module 210a might use a formula to determine the local control instructions to provide traffic light 230. To reduce pedestrian accidents, the remote instruction could include a heavier weight for pedestrian crossing variables in this example. Remote planning system 222, which may include different parameters than the local control systems 210a-210d, can be used to determine the most appropriate. The remote planning system 222 might determine that no cars are traveling east or south through an intersection for which the local controller system 210a is responsible. Remote planning system 222 can also receive local traffic data from local control system210 c, which may indicate that the northern arm’s left turn fraction is 1. The remote planning system may also receive local traffic data from the local controller system 210c. This may indicate that the left turn fraction for the northern arm is 1.

“In some cases, remote instructions may contain control instructions for traffic light 230. The remote instruction could include instructions to keep traffic light 230 on for four minutes, while there is a line at the intersection.

“The 300 process continues with stage F, in which the local controller system 210a makes a local decision on remote instruction and generates local controls instruction. The remote instruction may be ignored by the local control system, 210a when it generates the local control instruction. If there is low accident rate at an intersection, for example, the remote instruction may be ignored by the local control system.

“The 300th stage is completed with stage (G), where the local control instructions generated by local control system 210 a, and the remote instruction generated from remote planning system 222 are transmitted directly to the RTOS 232 which controls traffic light 230. To control the traffic light 230, the RTOS 232 uses both the remote instruction and the local control instruction to generate a signal for traffic control. If the generated traffic data indicates a situation that is not safe, the RTOS 232 can override the local instruction. If a police officer chases a suspect through a red stop sign at an intersection in a direction that is perpendicular with the direction for traffic light 230, the RTOS 232 could ignore remote and local instructions to change from red-to-green.

“Another example is that global traffic data might suggest that because there is heavy traffic heading north on a road, the green light phases at intersections in the northern direction should be extended. However, the RTOS 232 could determine that an emergency vehicle approaches the intersection from the east using local traffic data. The combined remote and local traffic data may allow the RTOS 232 to control the traffic light at the intersection for traffic going north. It will change phase from yellow to red to permit the emergency vehicle to cross.

The control operations described can be performed substantially in real time. The local control modules 210a-210d can provide traffic data at regular intervals. For example, they may deliver information at every 10 seconds, every 30 second, or at another interval. Remote planning system 222 may provide instructions for remote control, although they may not be provided at a set time. The remote planning system 222 can provide instructions for performing a specific pattern of light phase shifts over a period of time or until further instructions are given. The remote planning system 222 can also send instructions dynamically and/or simultaneously to any or all local control units 210a-210d in an area to deal with temporary or unusual conditions.

“In some cases, machine learning models (local and remote) may be substantially updated in real-time. The machine learning models may be updated periodically, for example, every 10 seconds, every 30 second, or at another time. Machine learning models can be updated when control instructions or data are received. This may or not be at a fixed time. Alternately, the machine-learning models can be updated dynamically or asynchronously in response to unusual or temporary conditions.

“In certain cases, local machine learning models might use federated learn. The system 200, for example, updates local machine learning models using weights that are determined by remote planning systems that receive data from multiple intersections. This provides system-level reinforcement learning via the network. To improve traffic flow, generalized rules may be applied to local control system by using the high-level view on traffic conditions.

“Additionally, different intersections may have different local control systems. This allows them to update their local machine learning models to target the most important variables at each intersection. Because each intersection can be controlled by distributed control, traffic flow control may be tailored to suit traffic conditions.

“The remote planning system can test traffic control techniques at multiple intersections. The remote planning system might reduce red light phases at half the intersections within a city. The remote planning system analyzes the changes and makes any necessary adjustments across multiple intersections. If reducing travel time by averaging fifteen minutes for all red light phases, the remote planning software may reverse the red light phase lengths to their original lengths. Remote planning systems may also update remote machine learning models to reflect results from testing. The remote planning system may automate the updating, testing, and analysis of remote machine learning models. The remote planning system might be given a list of variables to modify. The remote planning system might automatically identify variables that can be tested and then perform the tests.

It is possible to model traffic control at the system level and achieve high-level results which improve traffic conditions or assist authorities such as emergency responders. The system could be used to control traffic light phases at the remote planning level to coordinate traffic to clear intersections for a pileup, fire trucks racing to the scene, and police officers following a suspect in high-speed chase, among other things.

“FIG. “FIG. A system like the traffic light control system 200 could perform the process 400.

“A processing device that controls a traffic signal at an intersection receives sensor data (402) which provides images of the intersection. Some examples include data from cameras and/or sensors. Referring to FIG. 2 may receive camera data from cameras located near the traffic signal (230).

“The one or more processing units use one or several local machine learning models to identify the objects at the intersections and paths of objects based upon the sensor data (404). The local control system 210 might use the local machine-learning model 216 to identify vehicles, pedestrians and bicycles at intersections and their respective paths. Based on camera data from cameras located near the traffic light 230.

“The one or more processing units may transmit traffic data from local machine learning models to remote traffic planning systems over a network (406). The communication module 214 may be used by the local control system (210) to transmit the traffic data to remote planning system (222) over the network 220.

“The remote traffic planning system may send a remote instruction to one or more processing devices for the traffic signal that was determined by one or more remote machine-learning models (408). Remote planning system 222 could use remote machine learning model (224) to generate remote instructions. Remote planning system 222 could transmit the remote instruction via the network 220 to local control system210.

“The one or several processing devices may issue a control instruction to the intersection’s traffic signal. This instruction is determined using the remote instruction from remote traffic planning systems and the local instruction generated by one or more processing device (410). The local planning module 212 could generate a local control instructions based on traffic data received from cameras located near the intersection. The remote instruction, local instruction and generated traffic data may be provided by the local control system 210 to the RTOS 232. The RTOS 232 can analyze the remote instruction and local instruction to generate a traffic control instruction. The RTOS 232 could then control traffic light 230 by using the traffic control instruction. The local machine learning model (216) can generate control instructions that are sent directly to the traffic lights 230.

“Data processing apparatus” is a broad term. “Data processing apparatus” can refer to any type of apparatus, devices, or machines that process data. It could include a programmable processor, computer, multiple processors, or computers. An apparatus may include special purpose logic circuitry such as an FPGA (field-programmable gate array), or an ASIC. In addition to hardware, the apparatus may also contain code that creates an environment for the computer program, such as code that is processor firmware, protocol stack, an operating system or a combination thereof.

Click here to view the patent on Google Patents.