Real-time traffic management method and system

Abstract

The disclosure relates in particular to a real-time traffic management method and corresponding real-time traffic management system. The method comprises an initialization step during which ground sensors (GS) and video cameras (CAM), meant to detect vehicles on a road, are installed. Outputs of the ground sensors and of the video cameras are then fed into a neural network (NN) that is being trained on this basis. The method also comprises a real-time traffic management step during which the neural network detects vehicles in real time, and manages the traffic based on vehicles detections.

Description

[0001] REAL-TIME TRAFFIC MANAGEMENT METHOD AND SYSTEM

[0002] Traffic on road infrastructures tends to increase. This is true in most countries, and more specifically in developing countries. The growth of population generates increased use of such infrastructures, which often exceeds their expansion capacity. This in turn leads to issues such as traffic congestion, greater pollution, fast deterioration of the roads, delays or decrease of logistic productivity. Adaptive traffic management systems have been developed as one of key tool to mitigate such risks but still are expensive and challenging to implement. Data sensors may be deployed to provide input to traffic management methods implemented by traffic management systems. They may include ground sensors, radars, or video cameras, its combination.

[0003] Video recognition systems are becoming more and more common. They can be used for plate recognition, car classification, entry-exit systems management, etc. However, object detection and classification is challenging due, inter alia, to variations in images depending on viewpoints, scaling issues (a same object may appear with different sizes in images obtained from multiple video cameras or at different distances from a given video camera), or dynamic illumination conditions. Some tasks are particularly complex. Examples of complex tasks include dealing with overlapping objects, or determining the precise location of an object in a dynamic environment (with multiple moving objects around). Another key difficulty has to do with the fact that each camera typically undergoes a unique training process (typically based on artificial intelligence) for recognition purposes. A camera trained in a given country may be unable to deal with traffic in different countries. This may have to do with different driving habits, different levels of traffic, different equipment, etc. For example, driving on one side of the road (e.g. on the left side, as in the UK) or on the other (right side as in the USA) changes the analysis of road traffic. Different weather conditions may also have an impact. Snowy, rainy, or very dry and hot places may have to be treated differently. Places that are always sunny can be very different from places that are most often overcast. Places that have long nights during winter are not the same as those with more balanced durations between day and night, etc. The same is true of parameters such as the types of vehicles and the proportions of the respective types. For example, in certain cities, rickshaws may be very common, while in other places big SUVs are predominant, and in yet other areas, bicycles might be in very widespread use. Accordingly, collecting data in a certain place is not necessarily relevant for training a traffic management system for a different place. Other difficulties may involve the right to collect data (for example taking videos in public places may be considered a violation of citizens’ privacy in some regions, and be forbidden). Very fast networks may be needed due to the high throughput required for videos. Latency also matters. Video cameras are popular for traffic analytics. Traffic analytics applications normally have no real time requirements. In a traffic analytics application, the aim is typically to collect and process data in advance, for example to determine the intensity of traffic, quantity of vehicles, or types of the vehicles. Analytics so obtained can be used later when needed.

[0004] However, for real-time traffic management, things turn out to be more complex. Indeed, real-time traffic management requires an ability to assess a situation in real-time, and to take decisions in real-time. In particular, real-time traffic management may comprise controlling traffic lights to favor a road against another. As a rule of thumb, real-time is understood as an ability to react (assess a situation and take a decision) in less than 500 ms. According to a CCAM / C-ITS standard, vehicles can send CAM messages (Cooperative Awareness Message - safety related message with information about vehicle state, like position, direction or velocity) to a traffic management network and video algorithms must then process these videos in real time through RTSP Real Time Streaming Protocol) and synchronize both -data from video captured and CAM message attributes in terms of geospatial data. This is quite complex, and not suitable for situations where particular location (vehicle presence actuation) is pre-defined.

[0005] Cited as background of the invention, the document WO2018125508 A1 disclosed a real-time traffic management method. WO2018125508 A1 fails to teach how to combine sensor outputs of ground sensors and camera for feeding a neural network.

[0006] The invention seeks to improve the situation.

[0007] According to a first embodiment, a real-time traffic management method comprises an initialization step and a real-time traffic management step.

[0008] The initialization step comprises installing ground sensors to detect vehicles on a road. This may involve digging the road to install ground sensors beneath. Ground sensors outputs are timestamped. The initialization step comprises installing video cameras, in order to be able to take videos of said vehicles on said road at said position / location of specific ground sensor. Video cameras outputs are timestamped. Each video camera is associated with at least one respective ground sensor. Accordingly, it is possible to match data received from ground sensors and cameras by synchronizing them based on their timestamps. The initialization step comprises feeding the outputs of the ground sensors and of the video cameras, as well as their timestamps, into a neural network. This will be referred to as the “Reference Ground to Video” (or “RGtV” for short) method. RGtV may employ a machine learning process with data synchronization (timestamps on both the outputs from the video cameras and the outputs from the ground sensors). The training may rely on labeled datasets from reference ground sensors (it may be assumed with great confidence that the ground sensors provide reliable outputs). When a ground sensor states that a vehicle is present, this must be true, so if the neural network does not recognize a vehicle based on a corresponding camera image, it must be wrong and the neural network must adapt its weights to correct this error. Similarly, if the neural network reaches the conclusion that, based on camera images, there is a vehicle, while the ground sensor indicates that there is no vehicle, it is the ground sensor which must be correct. Vehicles detection based on the outputs of the ground sensors is used as a reference for supervised training of the neural network, to correct vehicles detection based on the outputs of the video cameras. In other words, for each measurement made by the ground sensor, and for each related measurement made (pictures taken) by a camera, a correlation (respective timestamps) makes it possible to identify which ground sensor measurements corresponds to which video camera measurement, and to verify if the inference made by the neural network on the basis of video cameras outputs is correct (in view of ground sensors outputs). Since the ground sensors are very reliable, while the analysis of video cameras outputs is complex and initially not reliable (until the training is completed), ground sensors are assumed to be correct and the neural network is configured to mimic ground sensors with video cameras. This is repeated until the difference between, on the one hand, the accuracy of vehicles detection based on the outputs of the video cameras and, on the other hand, the accuracy of vehicles detection based on the outputs of the ground sensors, remains below a set threshold. It may be assumed that the difference between these two accuracies is in fact substantially equal to the accuracy of the vehicles detection based on camera sensors (the accuracy of vehicles detection based on the outputs of the ground sensors being deemed perfect). In reality, it might be that the ground sensors sometimes provide erroneous outputs (exceptionally). However, this is so rare that it can be assumed to not exist, bearing in mind that the system does not necessarily have a way to verify whether the ground sensor was correct or not, while it can approximate the accuracy of the vehicle detection based on video cameras, by comparing its outputs to the outputs of the ground sensors, deemed to be exact. For example, if the video camera is able to provide results consistent with results obtained by the ground sensor in more than 99% of the cases (threshold = 100%-99% = 1%), the supervised learning may be deemed to have been completed.

[0009] Once the threshold has been crossed and therefore training is deemed to be sufficient, the real-time traffic management step comprises the neural network detecting vehicles in real time, based on the outputs of the video cameras and without access to ground sensors outputs (in particular the contemporaneous outputs). It further comprises managing the traffic based on vehicles detections. In other words, once the video cameras have been trained (on the basis of the ground sensors reference), it is possible to rely on video cameras only (not considering the ground sensors anymore) to manage the traffic, for example to control traffic lights or to display messages on electronic displays (dynamic road panels). More specifically, the output from the neural network may be communicated to a traffic management controller via a protocol known as the Multivalent Coordination Module protocol (MCM), with only a few parameters. This allows the trained RGtV model to be used not only for traffic analytics and forecasting, but also for real-time traffic management (traffic actuation mode), with low latency. The cameras can be trained in a fast, simple and reliable fashion thanks to the comparison with reference ground sensors.

[0010] According to a second embodiment, the ground sensors used in the real-time traffic management method according to the first embodiment are magnetometers. Magnetometers are particularly reliable compared to video cameras and may therefore provide a good reference point.

[0011] According to a third embodiment, the ground sensors used in the real-time traffic management method according to the first or second embodiment as well as the video cameras are geolocated, the outputs from the ground sensors and the video cameras being linked to their geolocation. Therefore, not only are the video cameras associated with corresponding ground sensors. Both are also linked by their geolocation, which enables a better correlation of the respective outputs from the two categories of sensors.

[0012] According to a fourth embodiment, video cameras used in the real-time traffic management method according to the first to third embodiments include a graphical processing unit (a.k.a. GPU) to preprocess the videos taken and assist in faster vehicles detection. Even with a GPU, a video camera is typically much cheaper than a magnetometer (considering both capex and opex). The GPU is useful to meet or exceed the performance and timing requirements of a real-time process, by transferring part of the computations normally carried out by backend servers to video cameras directly (namely, to their GPUs). On the other hand, GPUs tend to substantially increase the cost of video cameras. In addition, using an embedded GPU as in the fourth embodiment implies developing code (e.g. an integrated analytics engine) for such GPU, which code may be incompatible with other GPUs, possibly leading to a reduced vendor flexibility (increased dependence on the specific vendor concerned). With basic cameras (not equipped with GPUs), replacing a broken camera with another one is normally very simple (“plug and play”), while with more sophisticated cameras of the fourth embodiment, camera replacement may sometimes lead to traffic system performance disruptions (which may be due, for example, to increased configuration requirements - not plug and play). According to a fifth embodiment, the initialization step of the real-time traffic management method according to the first to fourth embodiments further comprises uninstalling ground sensors. Accordingly, when the initialization step completed and converged, the neural network being properly trained, it is possible to remove the ground sensors, or merely to ignores them. Physically removing ground sensors may be a costly operation, so alternatively the ground sensors may be kept in place, but the real-time traffic management method may simply ignore them. Indeed, the neural network having been trained on the basis of the video cameras outputs, it has become able to properly handle traffic management without input from the ground sensors.

[0013] According to a sixth embodiment, the initialization step of the real-time traffic management method according to the first to fifth embodiments further comprises installing additional video cameras, which additional video cameras are not associated with any ground sensors, and the real-time traffic management step further comprises the neural network detecting, in real time, vehicles in the vicinity of the additional video cameras, based on the outputs of the additional video cameras and without access to ground sensors outputs. Indeed, the neural network having been trained, it is able to handle additional video cameras in different locations without the burden of having to dig the road and embed ground sensors there. Preferably, the additional video cameras are installed in the same area as existing cameras or in an area with similar apparent characteristics, so that the neural network be as reliable as possible. Additional video cameras should not, without prior tests, be operated in locations exhibiting substantial differences with locations where the neural network was trained. For example, a location in a nearby country where traffic laws are different, or a location subject to very different weather conditions, or a location in an urban environment while the neural network was trained in the countryside (or vice versa), or a busy environment while the neural network was trained in a calm one (or vice versa).

[0014] According to a seventh embodiment, the initialization step of the real-time traffic management method according to the first to sixth embodiments further comprises installing different groups of ground sensors and video cameras, each group corresponding to a different geographic area, each group being analyzed by a separate neural network. This is advantageous as this makes it possible to adapt to various environments by dividing a geographical area into zones, each zone being covered by a respective group. It is possible to treat several separated zones in a single group, for example all villages in a geographical area may be treated the same, by a single neural network of a given group, while the countryside in between would be treated in a different group. In a possible embodiment, at least part of the real-time traffic management step can be enveloped as microservices and then wrapped in Docker, which is a set of platform as a service (PaaS) products that use OS- level virtualization to deliver software in packages called containers. Thanks to Docker, the RGtV method can be independently populated across traffic management or traffic analytics networks / systems. This provides a modular approach to configure the whole network. Accordingly, groups can be made more independent from one another. For example, software upgrade of microservices and hardware may be carried out in one group without affecting the other ones. This improves scalability across the system in which the method operates (network of video cameras, etc.). Relying on Kubernetes management orchestration is well adapted in this context.

[0015] According to an eighth embodiment, a computer program product comprises program instructions such that executing the program instructions implements the real-time traffic management step of a real-time traffic management method according to any of the first to seventh embodiments. The computer program may be written in performance efficient languages such as the C language which is convenient for real-time operations. If performance is an issue for certain components, relevant parts of the computer program may be written in assembly language, which may be even faster. If on the opposite, the hardware (e.g. servers, video cameras GPUs, etc.) is so numerous and fast that performance is not an issue, it is possible to write the computer program in higher level languages that are easier to learn and possibly easier to maintain, only the parts interfacing directly with hardware components (e.g. drivers) remaining in lower level languages such as C and assembly.

[0016] According to a ninth embodiment, a non-transitory computer readable storage device stores a computer program product according to the eighth embodiment. The non-transitory computer readable storage can be, for example, a purpose-built specialized NAS (Network- attached storage), or a RAID array attached to a server. While part of the computer program may be downloaded and executed on demand (e.g. with SaaS technologies), other parts (especially lower-level parts such as hardware drivers) are preferably downloaded once in the form of a setup software and executed locally, until a change is needed. For example, updates to the drivers may trigger the download of updated drivers, which are then executed locally (e.g. in video cameras or ground sensors), until yet another update comes in. According to a tenth embodiment, a real-time traffic management system comprises ground sensors arranged to detect vehicles on a road, ground sensors outputs being timestamped, and video cameras arranged to take videos of said vehicles on said road, said video cameras outputs being timestamped, each video camera being associated with at least one respective ground sensor. The real-time traffic management system is suitable for implementing the real-time traffic management method according to the first to seventh embodiments. It may comprise computer storage according to the ninth embodiment. The real-time traffic management system comprises an initialization module arranged to feed the outputs of the ground sensors and of the video cameras, as well as their timestamps, into a neural network, such that vehicles detection based on the outputs of the ground sensors is used as a reference for supervised training of the neural network, to correct vehicles detection based on the outputs of the video cameras, until the difference between, on the one hand, the accuracy of vehicles detection based on the outputs of the video cameras and, on the other hand, the accuracy of vehicles detection based on the outputs of the ground sensors, remains below a set threshold (for example, 1%, as was previously discussed). Once such a threshold has been reached, training is deemed completed. Optionally, training can be resumed if conditions change and justify an update. This could happen for example if road works change certain characteristics of the road (number of lanes, etc.), or different road signs are introduced, etc. But other than such exceptional circumstances, ground sensors have no reason to be used anymore and could even be removed, if practical.

[0017] The real-time traffic management module is arranged to have the neural network detect vehicles in real time, based on the outputs of the video cameras and without access to contemporaneous ground sensors outputs, and to manage the traffic based on vehicles detections.

[0018] According to an eleventh embodiment, the ground sensors of the real-time traffic management system according to the tenth embodiment are magnetometers. In general, these are the most appropriate ground sensors in the context of the invention. They are very reliable when it comes to detecting cars. In specific instances, other sensors can be added, as a complement to a magnetometer. For example, infrared beams can be sent from one position and detected at another position to determine that no vehicle is present in between (or on the contrary, that a vehicle is present, in case the beam is interrupted). Radar sensors could also be used, and any appropriate sensor may be considered. According to a twelfth embodiment, the ground sensors and the video cameras of the real-time traffic management system of the tenth or eleventh embodiments are geolocated. For example, they may store their GPS position. Typically, these sensors and cameras are fixed, and it is acceptable to store their GPS (or Glonas, or any other appropriate) position. In rare cases, such sensors or cameras may be moved around. For example, trial and error approaches may be undertaken to determine where it is best to locate them for optimal performance, until a specific location appears to stand out and be the best one. In such a scenario, it may make sense to have a GPS chip (or similar geolocation component) available in the cameras and / or ground sensors concerned, so that their geolocation be automatically available and up to date. The outputs from the ground sensors and the video cameras are linked to their geolocation. This facilitates learning by the neural network, since this provides further relevant information to interpret photos and videos taken by the cameras (being understood that a video can be considered a series of photos).

[0019] According to a thirteenth embodiment, the video cameras of the real-time traffic management system of any of the tenth to twelfth embodiments include a graphical processing unit to preprocess the videos taken and assist in faster vehicles detection. Not all cameras have to be equipped with GPUs, although this is a possibility. As was previously explained, having no GPU at all may well work, subject to the system otherwise having enough processing power to handle real-time operations.

[0020] According to a fourteenth embodiment, the initialization module of the real-time traffic management system of any of the tenth to thirteenth embodiments is further arranged to have its ground sensors uninstalled. When uninstallation is possible without too much work, this enables reusing the ground sensors to train the system for video cameras placed in other locations (by reinstalling the ground sensors there, until training is completed). Sometimes the work to remove a ground sensor may exceed the price of a new ground sensor, in which case uninstallation is economically not relevant, but may still be desired e.g. for ecological purposes.

[0021] According to a fifteenth embodiment, the initialization module of the real-time traffic management system of any of any of the tenth to fourteenth embodiments is further arranged to have additional video cameras installed, which additional video cameras are not associated with any ground sensors, and the real-time traffic management module is further arranged for the neural network to detect, in real time, vehicles in the vicinity of the additional video cameras, based on the outputs of the additional video cameras and without access to ground sensors outputs. This is based on the fact that once a supervised neural network has been sufficiently trained with inputs for which reference outputs are available (which enables corrections by the neural network when needed), such neural network is capable of providing proper outputs based on other inputs. The other inputs (from additional cameras) should be compatible with the training received. In other words, the additional cameras should be installed in locations which characteristics match the characteristics of the locations for which the training was carried out, as was already explained. Otherwise, consistent behavior of the neural network with the additional cameras should first be tested before the additional cameras are activated on the field for their final intended use (real-time traffic management).

[0022] According to a sixteenth embodiment, the initialization module of the real-time traffic management system of any of any of the tenth to fifteenth embodiments is arranged to have different groups of ground sensors and video cameras installed, each group corresponding to a different geographic area, each group being analyzed by a separate neural network. This is advantageous for reasons already stated in relation to the seventh embodiment.

[0023] The invention will be better understood by referring to the below drawings which show some specific examples of how it may be implemented.

[0024] Fig. 1 shows an infrastructure for real-time traffic management.

[0025] Fig. 2 is a diagram of a test bed according to the invention.

[0026] Fig. 3 is a workflow describing operations taking place in a test bed as shown in Fig. 2.

[0027] Fig. 4 shows, at a lower level, a container-based test bed architecture.

[0028] Fig. 5 illustrates a practical implementation derivable from the test bed of Fig. 2.

[0029] Fig. 1 is derived from Fig. 1 of A Knowledge Model for Automatic Configuration of Traffic Messages, Martin Molina, Monica Robledo, Department of Artificial Intelligence, Technical University of Madrid Campus de Montegancedo s / n, 28660 Boadilla del Monte (Madrid), Spain and Artificial Intelligence Group, E.S.C.E.T, University Rey Juan Carlos C / Tulipan s / n, 28993 Mostoles (Madrid), Spain, June 2001 . Fig. 1 schematically illustrates a system SYS to manage traffic in real time on a road RD (other roads are not shown, but the system is intended to manage traffic on a network of roads rather than on a single road). The system comprises a camera CAM, and a set of three ground sensors GS. The camera and the ground sensors are connected together via a network NET. The system SYS is controlled by a server SRV, also connected to the network NET, which comprises an artificial intelligence (neural network NN). One way the system SYS may manage traffic is by displaying indications for the vehicles drivers on a variable message panel VMP, on top of controlling traffic lights.

[0030] Fig. 2 is a diagram of a test bed according to the invention. The core of the diagram depicts a traffic light controller TLC. The TLC implements a Multivalent Coordination Module protocol MCM of the type that was described earlier. The traffic light controller TLC comprises two switches SW to interface respectively with ground sensors GS and with video cameras CAM.

[0031] The ground sensors of the test bed provide a reference data stream which may indicate, for each ground sensor, whether a respective vehicle is present or not at a given time and location, at a lower level. The ground sensors GS feed, via their switch SW, a traffic controller stack TCS.

[0032] Conventional video cameras may be used. They may be color or black & white. These cameras may be analog cameras. They may also be digital cameras (and may then use the IP protocol or any digital protocol suitable to exchange data over a network), and may output the videos in any appropriate format, such as H.264, MJPEG, MPEG2, MPEG4 or other, to provide a video stream that can be processed easily with off-the-shelf hardware and software complying with the selected format. Similar to the ground sensors, the cameras CAM feed, via their switch SW, the traffic controller stack TCS. But they do so via an artificial intelligence, because the output of video cameras, namely series of photographs encoded in a specific manner, tell nothing per se about what’s disclosed in these photographs, and requires interpretation. Therefore, the switch SW of the cameras sends either the raw videos (series of photographs) or (in case a camera has a GPU and is properly programmed to take advantage of it according to the invention), preprocessed video (to isolate presumably relevant features and assist the Al), via a fiber optics network FO_NET, to a traffic management control center TMCC. Fiber optic networks are advantageous due to their high bandwidth. However, if the videos (which require a lot of bandwidth) can fit within a network relying on a different technology, this is possible too.

[0033] The traffic management control center TMCC comprises a database DB with an application programming interface API to enable communication with other modules. The API is preferably a REST API, REST being the acronym for REpresentational State Transfer which designates an architectural style for distributed hypermedia systems (first proposed by Roy Fielding in year 2000). A Web API (or Web Service) conforming to the REST architectural style is an example of REST API. The traffic management control center TMCC also comprises a video server backend V SRV BE, which may be arranged to preprocess raw videos (for cameras not equipped with GPUs and proper supporting software), and which may also further analyze preprocessed video (either received from GPU equipped cameras or obtained locally by preprocessing raw videos received from basic cameras). The further analysis may rely on an artificial intelligence. The traffic management control center TMCC may, in addition to data processing, carry out various data orchestration steps such as data search, filtering and classification, analysis, logging, transaction processing, assembling, correlation and event control consistency between streams, statistics computation, transfer, representation and reporting, output analysis, dashboard etc. It can do so on separate server clusters. Several Al (which may use different deep learning models) may be combined to further improve the interpretation of the video camera events and improve the RGtV algorithm, to obtain a better traffic actuator hardware connected to traffic lights. Both traffic management plans and traffic analytics plans may be implemented in the traffic management control center TMCC. The traffic management control center TMCC may allow to vary the performance of the training by letting an operator of the system define this performance. Increase performance typically requires more training data and a longer training, and a lower threshold (for measuring and assessing the difference between the interpretation of the video streams collected by the video cameras and the outputs of the ground sensors, which should be minimal).

[0034] The traffic management control center TMCC may then feed, via the fiber optics network FO_NET, the MCM protocol of the traffic lights controller TLC. This MCM protocol may, in turn, feed the traffic controller stack TCS, which, based on its traffic algorithm TA, may control traffic lights TL.

[0035] The test bed is therefore configured for traffic management (and analytics, which is easier). It may do so with an RGtV algorithm set up as was previously discussed. The test bed combines two data sets of different natures (video and ground sensors such as magnetometer sensors) which it may synchronize both in time and in space (with timestamps and geolocation, as was discussed earlier). It uses these two sources to train an artificial intelligence (teaching it to improve its video recognition capacity) through reliable synchronized data received from ground sensors. This enables traffic actuation / management based on video streams only. Fig. 3 is a workflow describing operations taking place in a test bed as shown on Fig. 2. A zone of control ZC is monitored by a ground sensor GS and by a video camera CAM, driven respectively by a ground sensor driver GSD and by a video camera driver CAMD, which feed a timestamping and geolocation synchronization module TGS and an input database IN DB. Timestamping and geolocation synchronization may comprise reconciling timestamps provided by the ground sensor and by the video camera, thereby binding respective events (monitored at the same time by the ground sensor and the video camera), as well as reconciling geolocation information provided by the ground sensor and by the video camera, thereby similarly assisting in binding respective events.

[0036] Raw events recorded in the input database IN DB may be reviewed and fixed by a fixing unit FIX, arranged to analyze the flow of events that is received in the input database (data streams from the ground sensor and the video camera). The analysis may involve comparing the outputs of the ground sensor and of the camera, which in theory should be identical, but in practice are not (not always at least) - initially, the processing of the video camera outputs has to be tuned. Accordingly, the fixing unit FIX may continuously receive data from the ground sensor GS and camera CAM, and correct the way the video data from the camera is interpreted, with a neural network, based on an event model. The RGtV algorithm may be used. The neural network may implement open-source deep learning modules, such as the so-called “multiple linear regression” in combination with the GRU (gated recurrent unit) and LSTM (Long Short-Term Memory) specifically selected for traffic applications.

[0037] The fixing unit FIX may then feed an artificial intelligence, namely the training module TRN of a neural network. The training module adjusts the weights of the neural network based on information received from the fixing unit. If the fixing unit determines that the processing of the video camera outputs was successful in properly identifying the relevant event, i.e. it is consistent with the output of the ground sensor, then the weights do not require adjustment, otherwise the training module adjusts the weights to account for the inconsistency detected and try to get it right the next time. The training module TRN may be implemented in the form of a microservice (advantageous inter alia due to the lightweight and portable nature of containers implementing microservices). The microservice may be a Docker microservice. Alternatively, any other containerization technology that enables the packaging of an application (along with all its dependencies, if any) into a container to let such application run on any platform that supports that specific containerization technology could be used, in order to simplify the process of building, deploying, and running traffic management applications and making it easier to manage and scale them. The containerization technology (Docker or other) may be implemented on a server, or at the edge of the system. With containerization technologies, it is possible to create a microservice for training a neural network on how to process videos shot by the camera CAM in order to identify events consistent with the events detected by the ground sensor GS. The training module TRN may, once it is determined that the training has reached a sufficient quality (as has been discussed earlier - for example on the basis of a threshold), notify this fact to a Docker management entity DKR M (or any management entity for proper container technology). The Docker management entity DKR M may then release the training microservice, which has completed its intended task. The Docker management entity DKR M may then create a new microservice, to implement the tasks of a trained microservice (relying solely on the video camera, not on the ground sensor). To that end, the Docker management entity DKR M may: package the microservice (or group of microservices) into a container, create the Dockerfile specification to assemble an image of the application (which implements the trained microservice), and release a new container containing the trained microservice.

[0038] Fig. 4 shows, at a lower level, a container based test bed architecture. More specifically, Fig. 4 shows a Docker environment, in an initial state I DKR (on the left side) and later on, in an upgraded state U_DKR (on the right side). The environment shown comprises six Docker objects (numbered 1 to 6). Each of the six objects corresponds to a microservice for traffic control analysis, that has been uploaded into the environment. Initially, the Docker objects are ordinary microservices O_M and are shown in light grey. Once these microservices have been trained for traffic control or traffic analysis (e.g. with the “RGtV” method), they are upgraded into trained microservices T_M (shown in dark grey). If and whenever additional training is needed (for example based on an expansion of the statistical data sets at hand) a microservice may be synchronized though program patching or in cloud server in case of a remote application. Accordingly, when more accurate magnetometer outputs become available, they can be populated and used as a training basis for several microservices. This may improve the videos processing speed in the back end by better isolating basic attributes (such as presence / non presence, time, location) that can be determined and transferred via the MCM protocol to road controller for Traffic management purposes.

[0039] Thanks to Docker, each of the microservices may be managed in a uniform manner, and may operate independently of the particular characteristics of any proprietary server or edge hardware. This contributes to reliability and robustness of the system.

[0040] Fig. 5 illustrates a practical implementation that can be derived from the test bed of

[0041] Fig. 2. The architecture can be managed from a graphical user interface III. The architecture may rely on a parallel use of several microservices implementing artificial intelligences running together. The microservices shown (which are purely illustrative - other examples could be devised) each use a different model for their Al, and include (the microservices are identified by their learning model): moving average for traffic prediction (MS_MA), weighted moving average for traffic prediction (MS_WMA), multiple linear regression for traffic prediction (MS_MLR) and convolutional neural network for traffic prediction (MS_CNN). Other microservices are shown: a syslog microservice (MS_SL) dealing with logging of system events (for example for maintenance, security, or statistical purposes), and unspecified other microservices (MS_O). As can be seen, each of the microservices that deals with Al based processing of video data collected from cameras is associated with a respective database DB (which received video data). Each database may communicate with other entities (such as their microservice) via an application programming interface API (which is preferably a REST API). Another entity interfacing with other entities via an application programming interface API (that is preferably a REST API) is the elastic search ES. On the left side, three supplemental modules are shown:

[0042] • an adaptative selection module AS, which is designed to select, among available neural networks (MS_MA, MS_WMA, MS_MLR and MS_CNN), the one that is most promising; more elaborate schemes may also be implemented, where several neural networks are used in parallel and complement each other;

[0043] • a Docker module DKR, managing various containers hosting respective microservices;

[0044] • a system services module SS, which may manage tasks such as backup, database synchronization, monitoring of system behavior, load management (distributing processing across the system), or message queue management.

Claims

CLAIMS1 . A real-time traffic management method comprising: a. an initialization step comprising: i. installing ground sensors (GS) to detect vehicles on a road (RD), ground sensors outputs being timestamped, ii. installing video cameras (CAM) to take videos of said vehicles on said road, said video cameras outputs being timestamped, each video camera being associated with at least one respective ground sensor (GS), ill. feeding the outputs of the ground sensors and of the video cameras, as well as their timestamps, into a neural network (NN), so that vehicles detection based on the outputs of the ground sensors is used as a reference for supervised training of the neural network, to correct vehicles detection based on the outputs of the video cameras, until the difference between the accuracy of vehicles detection based on the outputs: of the video cameras and of the ground sensors remains below a set threshold; b. a real-time traffic management step comprising: i. the neural network (NN) detecting vehicles in real time, based on the outputs of the video cameras (CAM) and without access to ground sensors (GS) outputs, ii. managing the traffic based on vehicles detections.

2. The real-time traffic management method of claim 1 , wherein the ground sensors (GS) are magnetometers.

3. The real-time traffic management method of claim 1 or 2, wherein the ground sensors (GS) and the video cameras (CAM) are geolocated, the outputs from the ground sensors and the video cameras being linked to their geolocation.

4. The real-time traffic management method of any previous claim, wherein the video cameras (CAM) include a graphical processing unit to preprocess the videos taken and assist in faster vehicles detection.

5. The real-time traffic management method of any previous claim, wherein the initialization step further comprises: iv. uninstalling the ground sensors (GS).

6. The real-time traffic management method of any previous claim, wherein the initialization step further comprises: v. installing additional video cameras, which additional video cameras are not associated with any of the ground sensors (GS), and the real-time traffic management step further comprises: ill. the neural network (NN) detecting, in real time, vehicles in the vicinity of the additional video cameras, based on the outputs of the additional video cameras and without access to ground sensors outputs.

7. The real-time traffic management method of any previous claim comprising, during the initialization step, installing different groups of ground sensors and video cameras, each group corresponding to a different geographic area, each group being analyzed by a separate neural network.

8. A computer program product comprising program instructions such that executing the program instructions implements the real-time traffic management step of a real-time traffic management method according to any previous claim.

9. A non-transitory computer readable storage device storing a computer program product according to claim 8.

10. A real-time traffic management system (SYS) comprising: a. ground sensors (GS) arranged to detect vehicles on a road, ground sensors outputs being timestamped, b. video cameras (CAM) arranged to take videos of said vehicles on said road, said video cameras outputs being timestamped, each video camera being associated with at least one respective ground sensor (GS), c. an initialization module arranged to feed the outputs of the ground sensors and of the video cameras, as well as their timestamps, into a neural network (NN), so that vehicles detection based on the outputs of the ground sensors is used as a reference for supervised training of the neural network, to correct vehicles detection based on the outputs of the video cameras, until the difference between the accuracy of vehicles detection based on the outputs:of the video cameras (CAM) and of the ground sensors (GS) remains below a set threshold, d. a real-time traffic management module arranged to have the neural network (NN) detect vehicles in real time, based on the outputs of the video cameras and without access to ground sensors outputs, and to manage the traffic based on vehicles detections.1 1. The real-time traffic management system (SYS) of claim 10, wherein the ground sensors (GS) are magnetometers.

12. The real-time traffic management system (SYS) of claim 10 or 11 , wherein the ground sensors (GS) and the video cameras (CAM) are geolocated, the outputs from the ground sensors and the video cameras being linked to their geolocation.

13. The real-time traffic management system (SYS) of any of claims 10 to 12, wherein the video cameras (CAM) include a graphical processing unit to preprocess the videos taken and assist in faster vehicles detection.

14. The real-time traffic management system of any of claims 10 to 13, wherein the initialization module is further arranged to have its ground sensors (GS) uninstalled.

15. The real-time traffic management system of any of claims 10 to 14, wherein the initialization module is further arranged to have additional video cameras installed, which additional video cameras are not associated with any ground sensors (GS), and the real-time traffic management module is further arranged for the neural network (NN) to detect, in real time, vehicles in the vicinity of the additional video cameras, based on the outputs of the additional video cameras and without access to ground sensors outputs.

16. The real-time traffic management system of any of claims 10 to 15, wherein the initialization module is arranged to have different groups of ground sensors and video cameras installed, each group corresponding to a different geographic area, each group being analyzed by a separate neural network.

Images