Method and system for classifying vehicles by a data processing system

Abstract

In order to provide an efficient and reliable classification of vehicles by simple means, a method for classifying vehicles by a data processing system, in particular for classifying vehicles according to the characteristics of their drivers, is provided, the method comprising the steps of: collecting driving data for vehicles traveling in a predefined local area within a predefined time window, learning driving policies of one or more vehicles in the local area from said driving data, generating or using local predictors indicating predictions of definable driver behavior over a definable time range, sharing the local predictors with other vehicles in the local area to provide at least one composite predictor, redistributing the at least one composite predictor to vehicles in the local area, and locally classifying at least one of said vehicles into a definable vehicle class based on the at least one composite predictor and / or the local predictors to provide at least one local classification. Further, a corresponding system for classifying vehicles by a data processing system is provided.

Description

[Technical Field]

[0001] The present invention relates to a method and system for classifying vehicles by a data processing system. [Background technology]

[0002] Such methods and systems are known from the prior art and the corresponding prior art documents are listed below. [1] D. Petrovic, R. Mijailovic, D. Pesic, "Traffic Accidents with Autonomous Vehicles: Type of Collisions, Manoeuvres and Errors of Conventional Vehicles' Drivers", Elsevier Transportation Research Procedia, vol. 45, pp. 161-168, 2020. [2] Katherine Shaver, “Why your favorite bench might be there to thwart a terrorist attack,” The Washington Post, August 2018, [online], available at https: / / www.washingtonpost.com / local / trafficandcommuting / why-your-favorite-bench-might-be-there-to-thwart-a-terrorist-attack / 2018 / 08 / 27 / 28a863fc-9b49-11e8-b60b-1c897f17e185_story.html, accessed February 22, 2021 [3] L. Zanzi, A. Albanese, V. Sciancalepore, and X. Costa-Perez, "NSBchain: A Secure Blockchain Framework for Network Slicing Brokerage," ICC 2020~2020 IEEE International Conference on Communications, ICC, Dublin, Ireland, 2020, pp. 1–7, doi: 10.1109 / ICC40277.2020.9149414. [4] G. De Angelis, A. De Angelis, V. Pasku, A. Moschitta, and P. Carbone, "A simple magnetic signature vehicles detection and classification system for Smart Cities," 2016 IEEE International Symposium on Systems Engineering, / SSE, Edinburgh, 2016, pp. 1–6, doi: 10.1109 / SysEng.2016.7753170. [5] T. Moranduzzo, and F. Melgani, "Automatic Car Counting Method for Unmanned Aerial Vehicle Images," in IEEE Transactions on Geoscience and Remote Sensing, vol. 52, no. 3, pp. 1635–1647, March 2014, doi: 10.1109 / TGRS.2013.2253108. [6] Jiajun Zhu, David I. Ferguson, and Dmitri A. Dolgov, “System and method for predicting behaviors of detected objects through environment representation,” U.S. Patent No. 10,564,639 B1, granted February 2014. [7] Dougherty, John Anthony, Jordan Scott Burklund, Kristen Wagner Cerase, Stephen Marc Chaves, Ross Eric Kessler, Paul Daniel Martin, Daniel Warren Mellinger III, and Michael Joshua Shomin, "Methods And Systems For Managing Interactions Between Vehicles With Varying Levels Of Autonomy," U.S. Patent Application No. 16 / 727,179A, filed July 2, 2020.

[0003] The past few years have seen a great deal of hype surrounding the autonomous driving trend. As automakers develop self-driving SAE level 3+ systems to improve passenger safety and comfort, public and private transportation regulators (TRs) must establish new procedures to facilitate coexistence between autonomous and human-driven vehicles. For example, the increased vigilance of autonomous vehicles (AVs) near pedestrians represents a leading cause of accidents in which AVs are struck from behind after stopping to allow pedestrians the right of way (see [1]).

[0004] Modern urban architecture is evolving to counter terrorist threats. Instead of deploying bollards or concrete barriers, sidewalks and plazas are undergoing extensive redesign processes to maintain their comfort while making it possible to thwart terrorist vehicle ramming attacks (see [2]). Furthermore, high-risk urban areas, for example, central areas of interest for certain tourist destinations, may be restricted to AVs only, and therefore require a reliable system to identify and classify vehicles, as allowing access to the wrong vehicles would defeat the purpose of such protection solutions.

[0005] Related Work From the perspective of road systems, we are witnessing the emergence of a new security problem: the malfunction of AVs due to accidental or malicious tampering of their control units. Such malfunctions may result in a discrepancy between the driver's identity as expressed by the vehicle and the actual driver's identity. Currently, several technical solutions are available to improve the reliability of identity management, such as blockchain involving all the above parties or the use of Trusted Execution Environments (TEEs) on the vehicle side (see [3]). However, this scenario is so critical that traffic regulators (TRs) need a backup solution that can investigate, verify, or disprove the identity of vehicles circulating on roads, despite announcements to road management entities.

[0006] Automated traffic monitoring, including vehicle detection and tracking, has been widely studied in the past decades. Some literature utilizes sensors embedded in road infrastructure, e.g., buried induction loops, to design passive systems to assess the flow of vehicles passing through a specific area and determine vehicle models by comparing their electromagnetic fingerprint signatures with previously recorded references (see [4]). Another common approach is to use computer vision algorithms on footage generated by fixed cameras pointed at the road infrastructure, by unmanned aerial vehicles (UAVs) hovering over the area, or even by images coming from geostationary satellites (see [5]).

[0007] However, none of these methods can be applied to distinguishing between AVs and human-driven vehicles, which constitutes an open problem addressed by the present invention. The reason is that the two classes show significant differences only within isolated, sporadic traffic events, while their macro-movement patterns, such as following a straight centerline on a two-lane road, are very similar. Therefore, a classifier needs to observe the target vehicle over a large enough time window to capture this type of event. In particular, state-of-the-art solutions are based on fixed imagery, which is too slow to capture such movements in the case of roadside cameras or satellites, or on imagery generated by battery-powered UAVs, which are impractical and raise several efficiency concerns. Furthermore, integrating data generated by fixed road sensors, such as induction loops or cameras, is a viable option for tracking a target vehicle along its movement trajectory, but requires dense sensor deployment, which is uneconomical in areas with low traffic volume, such as rural areas.

[0008] The authors of [6] took a step in this direction by envisioning a method to predict the trajectory of an object in the environment. The AV determines the object's classification and state information, i.e., location, the lane the detected object is moving in, its speed, acceleration, whether it is entering the road, exiting the road, whether its headlights are on, its taillights are on, or whether its turn signals are on. However, the relationship between the object and the environment is not taken into account, since the features of the external environment are not part of the object's state.

[0009] Furthermore, [6] envisions a method for controlling an autonomous vehicle, i.e., adjusting its driving parameters based on the determined autonomy capability metrics of each identified vehicle. To this end, AVs can form clusters with one or more vehicles and share various information, such as the level of autonomy or speed, via vehicle-to-vehicle (V2V) communication. The system predicts the target's autonomous driving capability through observation of its external or non-external hardware devices or its driving behavior. This is performed by checking a set of features, such as the regularity of the vehicle's behavior, the degree to which nearby vehicles follow the center of the driving lane, the number of driving errors per unit time, compliance with local road and safety rules, the autonomous vehicle's reaction time, or its responsiveness. However, the proposed system includes a fixed number of test features, which do not necessarily capture the complexity of the target's driving behavior.

[0010] Furthermore, prior art reference U.S. Patent No. 2020 / 0207360A1 discloses a method for determining the autonomous capability metric (ACM) of a target vehicle, which involves determining the level of autonomy of the target vehicle, such as whether the vehicle is in fully autonomous mode, semi-autonomous mode, or manual mode, with the help of a vehicle autonomous driving system (VADS) component that collects data from various sensors present in the vehicle, such as cameras, radar, and LIDAR. The autonomous vehicles form clusters or caravans with one or more vehicles and share various information, such as the level of autonomy, speed, and velocity, with other vehicles in the caravan. The VADS component may be configured to detect the level of autonomy of the target vehicle with the help of various machine learning techniques or several prediction methods. The VADS component may adjust or modify the behavior models of other vehicles to more accurately reflect the determined level of autonomy of the target vehicle.

[0011] Furthermore, prior art document U.S. Pat. No. 8,660,734 B2 discloses a method for detecting external objects with the help of various types of sensors. A processor then analyzes the data and determines the classification and status of the target vehicle. The status of the target object can be determined with the help of its location, the lane the object is traveling in, its speed, acceleration, entering the road, exiting the road, and whether its headlights, taillights, or turn signals are on. This information can also be used to classify the target object. These observations and classifications can be achieved with the help of various types of machine learning techniques. The classification and status of the target vehicle can be shared with other nearby vehicles via a server, which can be considered a central server to which a group of cars are connected. [Prior art documents] [Patent documents]

[0012] [Patent Document 1] U.S. Patent No. 10,564,639B1 [Patent Document 2] U.S. Patent Application Serial No. 16 / 727,179A [Patent Document 3] US Patent No. 2020 / 0207360A1 [Patent Document 4] U.S. Patent No. 8660734B2 [Non-patent literature]

[0013] [Non-Patent Document 1] D. Petrovic, R. Mijailovic, D. Pesic, "Traffic Accidents with Autonomous Vehicles: Type of Collisions, Manoeuvres and Errors of Conventional Vehicles' Drivers", Elsevier Transportation Research Procedia, vol. 45, pp. 161-168, 2020. [Non-patent document 2] Katherine Shaver, "Why your favorite bench might be there to thwart a terrorist attack," The Washington Post, August 2018. [Non-patent document 3] L. Zanzi, A. Albanese, V. Sciancalepore, and X. Costa-Perez, "NSBchain: A Secure Blockchain Framework for Network Slicing Brokerage," ICC 2020~2020 IEEE International Conference on Communications, ICC, Dublin, Ireland, 2020, pp. 1-7, doi: 10.1109 / ICC40277.2020.9149414. [Non-patent document 4] G. De Angelis, A. De Angelis, V. Pasku, A. Moschitta, and P. Carbone, "A simple magnetic signature vehicles detection and classification system for Smart Cities," 2016 IEEE International Symposium on Systems Engineering, / SSE, Edinburgh, 2016, pp. 1–6, doi: 10.1109 / SysEng.2016.7753170. [Non-Patent Document 5] T. Moranduzzo, and F. Melgani, "Automatic Car Counting Method for Unmanned Aerial Vehicle Images," in IEEE Transactions on Geoscience and Remote Sensing, vol. 52, no. 3, pp. 1635–1647, March 2014, doi: 10.1109 / TGRS.2013.2253108. Summary of the Invention [Problem to be solved by the invention]

[0014] Traffic monitoring solutions developed and deployed over decades are not suited to the upcoming AV traffic. While there is no clear policy yet regulating the interaction between human-driven and autonomous vehicles, it is essential for traffic regulators to independently determine the nature of vehicle drivers with a high degree of confidence.

[0015] It is an object of the present invention to improve and further develop a method and system for classifying vehicles by means of a data processing system to provide an efficient and reliable classification of vehicles in a simple manner. [Means for solving the problem]

[0016] According to the present invention, the above-mentioned object is achieved by a method for classifying vehicles by means of a data processing system, in particular for classifying vehicles according to the characteristics of their drivers, the method comprising the following steps: - collecting driving data for a vehicle traveling within a predefined local area within a predefined time window; - learning a driving policy for one or more vehicles in the local area from the driving data; - generating or using a local predictor that indicates a prediction of a definable driver behavior over a definable time range; - sharing the local predictor with other vehicles in the local area to provide at least one composite predictor; - redistributing at least one composite predictor to vehicles within said local area; - locally classifying at least one of the vehicles into a definable vehicle class based on at least one composite predictor and / or local predictor to provide at least one local classification; Includes.

[0017] Furthermore, the aforementioned object is achieved by a system for classifying vehicles by means of a data processing system, in particular for classifying vehicles according to the characteristics of the driver of the vehicle, the system comprising: - collection means for collecting driving data relating to vehicles traveling within a predefined local area within a predefined time window; - learning means for learning a driving policy for one or more vehicles in the local area from the driving data; - generating or using means for generating or using a local predictor that indicates a prediction of a definable driver behavior over a definable time range; - sharing means for sharing the local predictor with other vehicles in said local area to provide at least one composite predictor; - redistribution means for redistributing at least one composite predictor to vehicles within said local area; classification means for locally classifying at least one of said vehicles into a definable vehicle class based on at least one composite predictor and / or local predictor to provide at least one local classification; Equipped with.

[0018] According to the present invention, it has been recognized that a local approach utilizing driving data for vehicles traveling within a predefined local area within a predefined time window can provide a highly efficient and reliable method. Furthermore, driving policies for one or more vehicles within the local area are learned from the driving data. Additionally, the method generates or uses local predictors that indicate predictions of definable driver behavior over a definable time range. The local predictors are then shared with other vehicles within the local area to provide at least one composite predictor for improved accuracy. After redistribution of the at least one composite predictor to vehicles within the local area, at least one of the vehicles is locally classified into a definable vehicle class based on the at least one composite predictor and / or local predictor to provide at least one local classification. The use of the at least one composite predictor and / or local predictor provides high accuracy in the classification of at least one of the vehicles. The proposed method and system provide high accuracy and low complexity in the detection and classification process.

[0019] Thus, in accordance with the present invention, an efficient and reliable classification of vehicles is provided by simple means.

[0020] According to one embodiment of the present invention, the vehicle class can provide information on whether the vehicle is driven autonomously or by a human. This information on the nature of the vehicle's driver is very important for many security issues and protection solutions in traffic regulation.

[0021] In further embodiments, the driving data can be collected from at least one sensor or on-board sensor of one or more vehicles, preferably one or more vehicles within a predefined local area, and / or from at least one road or environmental infrastructure sensor. This emphasizes that one or more suitable sensors can be provided on-board, in a vehicle other than the subject vehicle, and / or external to the vehicle. Alternatively or additionally, road or environmental infrastructure sensors can be used to provide the required driving data.

[0022] In further embodiments, the driving data may include abstract and / or synthetic data features. Depending on the particular application, any type of suitable data may be used within embodiments of the present invention.

[0023] According to a further embodiment, during the learning step, the unique implementation of the vehicle or autonomous vehicle can be preserved. No modifications or changes in the unique implementation are necessary.

[0024] In further embodiments, the local predictor can be provided as a locally adapted predictor adapted to the particular application situation of the method. More than one predictor can be generated or used. In general, one or more predictors can be tailored to two classification classes or vehicle classes.

[0025] According to a further embodiment, the classification step can be based on the predictor that provides the higher or highest accuracy score, which provides a highly accurate classification.

[0026] In a further embodiment, to provide even more enhanced accuracy of the classification, the local classifications can be shared with other vehicles, preferably to combine them, which also provides greater accuracy of the method.

[0027] In a further embodiment, the confidence estimates associated with the local classifications can be shared with other vehicles, preferably in order to combine them, which also provides greater accuracy of the method.

[0028] According to a further embodiment, one or more of said vehicles, i.e. one or more target vehicles, may be classified globally, preferably by combining the outputs of all local classifications and / or their associated confidence estimates. This feature provides a highly accurate and reliable classification of vehicles.

[0029] In further embodiments, at least one classification output can be transmitted to a traffic authority system, which provides highly reliable enforcement of appropriate automated control policies based on, for example, vehicle type.

[0030] In further embodiments, the method can be performed in one or more vehicles and / or in one or more external or edge data processing systems. The method can be performed in autonomous vehicles and / or human-driven vehicles. Also, corresponding systems can be provided on autonomous vehicles and / or human-driven vehicles. Alternatively or additionally, sections of the method and / or components of the system and / or data processing system can be performed at the network edge and can be supplied with driving data.

[0031] According to a further embodiment, the method can be implemented as a machine learning method, preferably in an edge computing (EC) network with edge computing servers. Such implementation of the method and system may depend on the particular application. The machine learning method provides a highly efficient and reliable classification of vehicles and allows for permanent and continuous updating of the method due to changing traffic conditions.

[0032] In further embodiments, the method can be performed using a computing server, preferably an edge computing server, communicating via a direct link, via a cloud backend, and / or via a connected, cooperative automated mobility (CCAM) platform. The type of communication can be selected to optimize accuracy in vehicle classification.

[0033] According to a further embodiment, the method may include having the vehicle train the neural network and update the weights on an assigned server or edge computing server, and a distributed learning mechanism across the network provided by the server may be implemented to update the global model within the network.

[0034] Advantages and aspects of embodiments of the present invention are listed below.

[0035] According to embodiments of the present invention, autonomous vehicles can be detected by predicting driving characteristics and classifying them based on their respective prediction scores.

[0036] In a further embodiment, the classification output can be sent to a traffic authority system to enforce appropriate automatic control policies based on the type of vehicle.

[0037] Within the scope of the present invention, a method for automatically detecting whether a vehicle is being driven autonomously or by a human may be provided, which may include one or more of the following steps: 1) Learning a driving policy for an autonomous vehicle according to abstract data features and / or driving data collected in a certain time window, thereby preserving the unique implementation of the autonomous vehicle or driving unit. 2) Sharing at least one or more local predictions within the vehicle pool provided by the vehicles in order to combine them for better accuracy. 3) Redistributing the combined, potentially better predictors to vehicles in the vehicle pool. 4) Locally classify the target vehicle based on the predictor that provides a higher accuracy score. 5) Sharing local classifications and / or associated confidence estimates within a vehicle pool in order to combine them for better accuracy. 6) Globally classify the target vehicle by combining all local classification outputs and / or their confidence estimates in the vehicle pool.

[0038] In contrast to the system of [6], for example, embodiments of the inventive solution do not rely on explicit assessment of key performance indicators (KPIs) of target vehicles, but rather allow driving behavior to be inferred natively by ad-hoc predictors, as discussed later in this document. Furthermore, embodiments of the inventive solution can enable a distributed learning approach in which all AVs in an area governed by the same TR can participate in training predictors and classifiers of a global model.

[0039] According to further embodiments of the present invention, a system can be provided that automatically determines whether a vehicle on a road is an autonomous or human-driven entity by utilizing on-board sensors of other vehicles traveling in the same area and / or road infrastructure sensors, thereby building a classifier based on locally adapted predictors.

[0040] One embodiment of the present invention provides an independent backup system that i) can verify the vehicle AV / human driver's announcements and ii) can minimize the data scarcity issues of state-of-the-art solutions in sparsely populated areas.

[0041] There are several ways of advantageously designing and further developing the teachings of the present invention, and for this purpose, reference should be made to the following description of examples of embodiments of the invention that are illustrated in the drawings. [Brief explanation of the drawings]

[0042] [Figure 1] FIG. 1 illustrates the building blocks and respective implementation entities of one embodiment of the present invention. [Figure 2] FIG. 1 illustrates an edge computing scenario according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0043] Embodiments of the present invention overcome the limitations of the prior art discussed above by taking advantage of the increasing prevalence of connected vehicles with sensors on roads and in infrastructure by leveraging large amounts of locally acquired data.

[0044] One embodiment of the present invention includes services running on both classes of vehicles, for example but not limited to, also running at the network edge and fed by vehicle sensor data, that perform two main tasks described below. Furthermore, it brings a new approach to classification based on the performance of predictors tailored to the two classification classes, e.g., autonomous vehicles or human-driven vehicles.

[0045] Training a local model constitutes the first task. Such a model aims to learn the driving policy of the vehicle's driver, i.e., predict the driver's behavior in terms of steering wheel angle, throttle, and / or braking, and / or driving accuracy, based on a given window of collected sensor data. The sensor data includes a rich set of features, such as the vehicle's current speed and acceleration vectors, as well as past trajectories acquired by, for example, a Global Satellite Navigation System (GNSS), cameras, or ambient radar / LiDAR images. It is worth noting that the list of features can be directly adapted based on the pool of available onboard and infrastructure sensors.

[0046] Thus, the local predictor displays or outputs a prediction of the driver's behavior over a particular time range, with the past time window and future time period adjusted to minimize the prediction error according to the literature relevant to the particular implementation of the predictor.

[0047] Learning an AV's driving policy by observing driving behavior based on synthesized data features is beneficial because it i) likely avoids the complexity of the autonomous driving model, and ii) preserves the proper implementation of the autonomous driving unit, which does not need to be disclosed, thus encouraging automakers to implement the system. Nevertheless, even if it cannot achieve the accuracy required to perform the autonomous driving task, the predictor can still provide sufficient performance for final classification purposes.

[0048] Once the training phase is complete, the adjusted predictors are shared with the traffic regulator TR, which collects predictors tailored to both classes. The TR then combines the obtained predictors to build several finer predictors by combining the predictors at its disposal. Note that after the first training phase, the vehicle continues to train its predictors in an online fashion.

[0049] The final predictor is then shared with the vehicles in a federated-like fashion. Now, each vehicle can perform a second task, i.e., apply both predictors to any of the surrounding vehicles, since most of the input feature set can be derived from their macroscopic behavior. Missing features, e.g., camera images of the target vehicle, can be addressed by encoding techniques or generative methods, which are combined in the adaptation layer of a model, e.g., a pre-trained generative adversarial network model, to generate the missing images from the target vehicle's perspective.

[0050] By looking at the predictive performance of both predictors and recognizing which type of vehicle, AV or human-driven, they were trained on, the present invention classifies the target vehicle according to the predictor class that provides the highest score. The high-level building blocks of the present invention are shown in Figure 1, along with the entity that performs each function: vehicle or TR. After collecting onboard and infrastructure sensor data, each vehicle trains a local predictor, which is then shared with the TR, which is responsible for combining all received predictors and distributing the final predictor to the vehicle. This process continues in a closed loop, thereby improving the predictor accuracy by increasing the number of observed samples and associated vehicles over time. Upon receiving the improved predictor, the vehicle performs the target classification and returns its classification output, with an enhanced confidence measure, to the TR, which can finally combine all classification outputs to derive the vehicle class and overall confidence.

[0051] In the following, we present several embodiments of the present invention obtained through machine learning techniques in an edge computing EC network scenario.

[0052] Embodiment 1 In this embodiment, we present a possible implementation of the present invention in a scenario with edge computing EC servers, as shown in Figure 2. In particular, we envision a multi-server architecture in which the EC servers can communicate via direct links or via a cloud backend and / or a connected cooperative automated driving platform CCAM.

[0053] Since the neural network model can be easily distributed by sharing the weights of the links between neurons among all federation parties, all vehicles train the same neural network model and update the weights in their assigned EC servers. Note that a distributed learning mechanism among EC servers can be introduced to update the global model in the network.

[0054] After the training phase, the final predictors are returned to the vehicles, which apply them to their target vehicles by obtaining the required target feature sets through a generative model, e.g., a generative adversarial network model.

[0055] By selecting the predictor that provides the highest accuracy, each vehicle obtains a classification for the target vehicle. Each time a vehicle obtains a classification output for any target vehicle, it sends it to its EC server, which collects all classifications and evaluates their confidence based on the number of vehicles that provided classification outputs for the same target, or how many vehicles saw the target and for how long. Finally, one or more EC servers output the labels and their respective confidence levels for the TR to check the legitimacy of all vehicles on the road.

[0056] It should be noted that embodiments of this system do not necessarily need to be offline, as continuous learning techniques applied to both the vehicle and the EC server can provide continuous improvements to the predictors and classifiers.

[0057] Owners of the CCAM platform (traffic regulatory entities, automobile manufacturers, and / or mobile operators) may rely on AV classification to enhance road safety by instructing traffic authorities and drivers who subscribe to their services to prevent accidents.

[0058] Embodiment 2 In this embodiment, we utilize the upcoming network slicing paradigm and assume that the automaker holds a network slice within some service operator network that connects all of its vehicles. Since the automaker is primarily concerned with customer safety, it can implement this invention to classify vehicles from other vendors and adjust the on-board advanced driver-assistance system (ADAS) based on their characteristics.

[0059] Specifically, under these operating conditions, vehicles from one manufacturer can share their prediction models to build a few accurate predictors, which can then be distributed to vehicles. Again, such predictors can be used for vehicle classification.

[0060] It should be noted that the accuracy of this approach increases with the amount of data available for training the predictor as well as the number of vehicles providing classification outputs for a particular subject.

[0061] Many modifications and other embodiments of the inventions described herein will come to mind to one skilled in the art to which these inventions pertain having the benefit of the teachings presented in the foregoing descriptions and the associated drawings. Therefore, it is to be understood that the invention is not to be limited to the specific embodiments disclosed, and that modifications and other embodiments are intended to be included within the scope of the appended claims. Although specific terms are employed herein, they are used in a generic and descriptive sense only and not for purposes of limitation.

Claims

1. 1. A method for classifying vehicles by a data processing system, in particular for classifying vehicles according to the characteristics of their drivers, comprising: - collecting driving data for a plurality of vehicles traveling within a predefined local area within a predefined time window; - learning a driving policy for one or more vehicles in the local area from the driving data; - generating or using a local predictor that indicates a prediction of a definable driver behavior over a definable time range; - sharing the local predictors with other vehicles in the local area to provide at least one composite predictor, wherein the sharing of the local predictors comprises transmitting local predictors from a plurality of vehicles to a central entity, and the at least one composite predictor is generated by combining the local predictors using a federated learning scheme; - distributing the at least one composite predictor to vehicles within the local area; - locally classifying at least one of the vehicles into a definable vehicle class based on the at least one composite predictor and / or the local predictor to provide at least one local classification; wherein the vehicle class provides information about whether the vehicle is driven autonomously or by a human.

2. 2. The method of claim 1, wherein the driving data is collected from at least one sensor or on-board sensor of one or more vehicles, preferably one or more vehicles within the predefined local area, and / or from at least one sensor of road or environmental infrastructure.

3. The method of claim 1 or 2, wherein the driving data includes abstract data features and / or synthetic data features.

4. 4. The method of claim 1, wherein during the learning step, a unique implementation of the vehicle or autonomous vehicle is saved.

5. 5. The method of claim 1, wherein the classification step is based on the predictor that provides the higher or highest accuracy score.

6. 6. The method according to any one of claims 1 to 5, wherein the local classifications are shared with other vehicles, preferably in order to combine them.

7. 7. The method of any one of claims 1 to 6, wherein confidence estimates associated with local classifications are shared with other vehicles, preferably for combining them.

8. 8. The method of any one of claims 1 to 7, wherein one or more of the vehicles, i.e. one or more target vehicles, are classified globally, preferably by combining the outputs of all local classifications and / or their associated confidence estimates.

9. 9. The method of claim 1, wherein at least one classification output is transmitted to a traffic authority system.

10. 10. The method of claim 1, wherein the method is performed in one or more vehicles and / or in one or more external or edge data processing systems.

11. The method according to any one of claims 1 to 10, wherein the method is preferably implemented as a machine learning method in an edge computing EC network having edge computing servers.

12. 12. The method according to claim 1, wherein the method is performed using a computing server, preferably an edge computing server, communicating via a direct link, via a cloud backend and / or via a connected collaborative automated driving platform CCAM.

13. 13. The method of claim 1, wherein the vehicle trains a neural network and updates weights on an assigned server or an edge computing server.

14. A system for classifying vehicles by means of a data processing system, in particular for classifying vehicles according to the characteristics of their drivers, preferably for carrying out the method according to any one of claims 1 to 13, - collection means for collecting driving data for a plurality of vehicles traveling within a predefined local area within a predefined time window; - learning means for learning a driving policy for one or more vehicles in the local area from the driving data; - generating or using means for generating or using a local predictor that indicates a prediction of a definable driver behavior over a definable time range; - sharing means for sharing the local predictors with other vehicles in the local area to provide at least one composite predictor, wherein the step of sharing the local predictors comprises transmitting local predictors from a plurality of vehicles to a central entity, and the at least one composite predictor is generated by combining the local predictors using a federated learning scheme; and - distribution means for distributing said at least one composite predictor to vehicles within said local area; classification means for locally classifying at least one of the vehicles into a definable vehicle class based on the at least one composite predictor and / or the local predictors to provide at least one local classification; wherein the vehicle class provides information on whether the vehicle is driven autonomously or by a human. system.

Images