Method for generating training data for training an AI-based behavior planner for a partially automated vehicle

Abstract

According to various embodiments, a method for generating training data for training an AI-based behavior planner for an at least partially automated vehicle is provided, comprising providing vehicle trajectories in traffic scenes without an accident event, providing accident event data elements, each describing an accident event, training a machine learning model based on the provided vehicle trajectories without accident events to reconstruct the vehicle trajectories without accident events, and generating the training data for training the behavior planner by, for each accident event data element,Generating a trajectory of at least one vehicle involved in the accident using the machine learning model and creating a respective training data element from the training data to train the behavior planner from the generated trajectory of the at least one vehicle involved in the accident and information that the trajectory of the vehicle involved in the accident belongs to a critical traffic scene.

Description

[0001] The present disclosure relates to methods for generating training data for training an AI-based behavior planner for an at least partially automated vehicle.

[0002] When controlling at least partially automated vehicles, it is advisable to assess the criticality of each traffic scenario and adjust the control accordingly. For example, if a traffic scenario is very dangerous, an autonomous vehicle should drive more cautiously. A machine learning model (e.g., a neural network) can be used to assess a traffic scenario. This model is trained accordingly and forms the basis for an AI-based behavior planner. However, this requires comprehensive training data containing examples of both critical and non-critical traffic scenarios. Therefore, approaches for generating training data for training an AI-based behavior planner for at least partially automated vehicles are desirable.

[0003] According to various embodiments, a method for generating training data for training an AI-based behavior planner for an at least partially automated vehicle is provided, comprising providing vehicle trajectories in traffic scenes without an accident event, providing accident event data elements, each describing an accident event, training a machine learning model based on the provided vehicle trajectories without accident events to reconstruct the vehicle trajectories without accident events, and generating the training data for training the behavior planner by providing, for each accident event data element, • Generating a trajectory for at least one vehicle involved in the accident event using the machine learning model and • Forming a respective training data element from the training data to train the behavior planner from the generated trajectory of the at least one vehicle involved in the accident and information that the trajectory of the vehicle involved in the accident belongs to a critical traffic scene.

[0004] The method described above enables the automated generation of a training dataset (with the labels necessary for training) for an AI-based behavior planner for evaluating traffic scenes, which contains realistic trajectories for both traffic scenes without an accident and traffic scenes without an accident, and therefore enables the training of the behavior planner so that it correctly evaluates traffic scenes occurring in practice.

[0005] The following are various examples of implementation.

[0006] Exemplary embodiment 1 is a method for generating training data as described above.

[0007] Exemplary embodiment 2 is the method according to exemplary embodiment 1, wherein the respective training data element of the training data for training the behavior planner is further formed from information that the trajectory of the vehicle involved in the accident event belongs to a critical traffic scene.

[0008] This enables supervised training of the behavior planner.

[0009] Embodiment 3 is the method according to embodiment 2, wherein the machine learning model is trained to reconstruct the vehicle trajectories without accident events from their endpoints and information about the areas in which they extend.

[0010] This then allows for a transfer to traffic scenarios with accident events, since the endpoint (location of the accident) and uncertainty limits are usually known for these.

[0011] Embodiment 4 is the method according to embodiment 3, wherein the machine learning model is trained such that, for a training input which specifies the endpoint of a vehicle trajectory and an area for a course of the vehicle trajectory before the endpoint, it outputs a vehicle trajectory which lies within the area.

[0012] The machine learning model can therefore be forced to strictly adhere to a defined range for the trajectory. This ensures that a trajectory generated using the machine learning model is no worse (e.g., less realistic) than a trajectory generated conventionally (provided the specified range corresponds to the uncertainty range of the conventional generation method).

[0013] Embodiment 5 is a method according to embodiment 3 or 4, wherein each accident event data element contains a reconstructed trajectory with associated uncertainty bounds and, for each training data element for the machine learning model, the range for the course of each of the at least one vehicle trajectory is determined by uncertainty bounds of a respective reconstructed trajectory whose endpoint is placed on the endpoint of the vehicle trajectory.

[0014] In this way, areas for the trajectories are determined that lie within the plausible range (according to the reconstruction rules).

[0015] Exemplary embodiment 6 is the method according to one of the exemplary embodiments 3 to 5, wherein for each accident event data element the trajectory of the at least one vehicle involved in the accident event is generated by determining an area for the course of the trajectory of the at least one vehicle involved in the accident event and generating the trajectory of the at least one vehicle involved in the accident event by supplying the determined area and the accident location as the endpoint of the trajectory of the at least one vehicle involved in the accident event to the machine learning model.

[0016] This allows realistic trajectories to be generated for traffic scenes involving accidents.

[0017] Embodiment 7 is the method according to one of embodiments 1 to 6, further comprising generating additional training data elements for training the behavior planner, wherein each additional training data element is formed from at least one of the vehicle trajectories in traffic scenes without an accident or a trajectory generated by the machine learning model for a traffic scene without an accident and information that the at least one trajectory of the vehicle involved in the accident belongs to a non-critical traffic scene.

[0018] This means that the training data for the behavior planner includes training examples for critical and non-critical traffic scenarios.

[0019] Exemplary embodiment 8 is the method according to one of exemplary embodiments 1 to 7, wherein the training data elements of the training data for training the behavior planner are supplemented with position information about further objects in the respective traffic scene and / or the information for one or more further objects that they are not visible in the respective traffic scene as input for the behavior planner.

[0020] This allows a more complete picture of the respective traffic scene (e.g. from the perspective of an ego-vehicle) to be generated for the training data elements of the behavior planner, which the behavior planner works with.

[0021] Exemplary embodiment 9 is the method according to one of the exemplary embodiments 1 to 8, wherein information about objects that are not visible in the respective traffic scene is removed from the training data elements of the training data for training the behavior planner.

[0022] This means that the input received by the behavior planner contains information that it can realistically contain in practice.

[0023] Exemplary embodiment 10 is a data processing device which is configured to carry out a method according to one of the exemplary embodiments 1 to 9.

[0024] Embodiment 11 is a computer program with instructions which, when executed by a processor, cause the processor to perform a method according to one of embodiments 1 to 9.

[0025] Embodiment 12 is a computer-readable medium that stores instructions which, when executed by a processor, cause the processor to perform a method according to one of embodiments 1 to 9.

[0026] In the drawings, similar reference numerals generally refer to the same parts in all the different views. The drawings are not necessarily to scale, with the emphasis generally placed on illustrating the principles of the invention. Various aspects are described in the following description with reference to the drawings. Fig. Figure 1 shows a vehicle. Fig. Figure 2 shows an example of an accident at an accident site with a reconstructed trajectory for a first vehicle and a reconstructed trajectory for a second vehicle. Fig. Figure 3 shows another symbolic example of a possible reconstructed dynamic up to a collision point. Fig. Figure 4 illustrates a process for generating a training dataset based on a dataset of recorded trajectories and a dataset in which accidents were recorded and reconstructed. Fig. Figure 5 shows a flowchart illustrating a method for generating training data for training an AI-based behavior planner for at least a partially automated vehicle according to one embodiment.

[0027] The following detailed description refers to the accompanying drawings, which illustrate specific details and aspects of this disclosure in which the invention can be implemented. Other aspects may be used, and structural, logical, and electrical modifications may be made without deviating from the scope of the invention. The various aspects of this disclosure are not necessarily mutually exclusive, as some aspects of this disclosure may be combined with one or more other aspects of this disclosure to form new aspects.

[0028] Several examples are described in more detail below.

[0029] Fig. Figure 1 shows vehicle 101.

[0030] In the example of Fig. 1 is a vehicle 101, for example a motor vehicle such as a car or truck, equipped with a vehicle control device (e.g. an Electronic Control Unit (ECU)) 102.

[0031] The vehicle control unit 102 has data processing components, e.g., a processor (e.g., a CPU (central processing unit)) 103 and a memory 104 for storing control software 107, according to which the vehicle control unit 102 operates, and data that is processed by the processor 103. The processor 103 executes the control software 107.

[0032] For example, the stored control software (computer program) contains instructions which, when executed by the processor, cause the processor to perform 103 driver assistance functions or even control the vehicle autonomously.

[0033] The control software 107 is transferred, for example, from a computer system 105 to the vehicle 101, e.g., via a network 106 (or also with the help of a storage medium such as a memory card). This can also happen during operation (or at least when the vehicle 101 is with the user), since the control software 107 is updated to new versions over time.

[0034] The tax software 107 can, for example, be trained using machine learning (ML). This means that the tax software 107 implements one or more ML models 108 (or "machine learning models") that are trained on the basis of training data, in this example from the computer system 105. The computer system 105 thus implements an ML training algorithm for training the one or more ML models 108. For example, the tax software 107 implements a (Kl-based) behavior planner that is based on the one or more machine learning models 108.

[0035] The control software 107 determines control actions for the vehicle (such as steering, braking, etc.) from input data 109 available to it, which contains information about the environment or from which it derives information about the environment (such as by detecting other road users, e.g., other vehicles). This includes, for example, sensor data such as from a camera on the vehicle or information obtained through communication with other vehicles or devices at the roadside.

[0036] In road traffic, it frequently happens that a vehicle finds itself in an unfamiliar traffic situation that was not anticipated during the design phase of the control software, and especially during the training phase of the one or more ML models. Human drivers typically have an intuitive sense based on their experience, which leads them to drive particularly cautiously in such situations. Automated vehicles (i.e., control software), on the other hand, typically lack such a mechanism.

[0037] To train an automated vehicle, i.e., the control software 107 or one or more ML models 108, with such empirical data, corresponding knowledge information in the form of datasets is required, which represent the empirical data from real-world operation. Typically, datasets are available containing... (1) trajectories recorded in real road traffic from, for example, drone data, but also (2) retrospectively collected and reconstructed accidents.

[0038] Datasets containing real trajectories (e.g., recorded using drones) generally do not record accidents, but the trajectories usually have a very high accuracy.

[0039] In contrast, retrospectively investigated accidents record all available information, which serves as the basis for reconstructing the sequence of events leading up to the collision. In these cases, information from the driving process itself is generally unavailable, making a reliable reconstruction of the sequence of events impossible. This gap could be filled by additional information from so-called event data recorders (EDRs).

[0040] As a result, the accident scenarios in accident databases are inherently uncertain and contain reconstruction artifacts that are fundamentally unavoidable. Reconstruction artifacts include, for example, simplifications regarding the trajectory that led to the collision (e.g., the assumption that the vehicle was traveling in the middle of the lane) or the dynamics (constant speed and / or constant acceleration).

[0041] Fig. Figure 2 shows an example of an accident from an accident database at an accident site 201 with a reconstructed trajectory 202 for a first vehicle 203 and a reconstructed trajectory 204 for a second vehicle 205 (and two obstacles 206, 207). The trajectory reconstruction includes (in addition to the spatial path) the respective speed profile 208 (first vehicle), 209 (second vehicle) and acceleration profile 210 (first vehicle), 211 (second vehicle).

[0042] Fig. Figure 3 shows another symbolic example of a possible reconstructed dynamic (speed profile) in the form of a dashed line 301 up to a collision location (or collision time) 300.

[0043] As in the examples of Fig. 2 and Fig. As illustrated in Figure 3, the reconstructed trajectories (especially the reconstructed dynamics) are very schematic and, at least in many cases, do not correspond to realistic trajectories. A machine learning model (e.g., one of the models in Figure 108, such as a behavior planner) could be better trained to correctly assess realistic traffic scenarios using more realistic trajectories, i.e., with such unrealistic trajectories as training examples.

[0044] To provide more realistic training data for better training, a method is provided according to various embodiments which uses machine learning methods to link traffic observation data - i.e., "accident-free" driving - and accident data.

[0045] This method can use machine learning to generate training examples for critical traffic scenarios in which no reconstruction artifacts are present, or at least their presence is significantly reduced. Together with training examples from normal driving scenarios (i.e., non-critical traffic scenarios), this can create a training dataset of training examples (training data elements), where each training example, for example, is an input data element representing a traffic scenario (including one or more trajectories for an ego vehicle and one or more other road users) and contains ground truth data, such as a label indicating whether the traffic scenario resulted in an accident.

[0046] The provided method enables the automated processing of existing data, which includes both critical traffic scenarios from reconstructed accident data and non-critical traffic scenarios (normal driving, e.g., from drone data or data from stationary traffic observation), and the automatic generation of training examples including labels as described above. The resulting training examples can then be used to train a machine learning model (e.g., one of the ML models 108, such as one or more neural networks, e.g., a behavior planner) that replicates human driving knowledge (resulting from experience) (e.g., outputs a criticality measure for input sensor data representing a traffic scenario) and is used, for example, in an automated vehicle, enabling it to react anticipatorily to unknown or critical traffic scenarios and to mitigate developing critical traffic situations.

[0047] An embodiment of the provided method is described below with reference to Fig. 4 described.

[0048] Fig. Figure 4 illustrates a process for generating a training dataset based on a dataset of recorded trajectories and an (accident) dataset 405 in which accidents were recorded and reconstructed.

[0049] As described above, the training dataset that is generated contains training examples for critical traffic scenes and training examples for non-critical traffic scenes, which are automatically labelled accordingly (i.e., the training examples that are generated contain corresponding labels, e.g., "critical" and "non-critical").

[0050] The process includes the following procedural steps: 1) Preprocessing the dataset with recorded trajectories: First, the dataset with recorded trajectories is filtered from observational data containing the recorded trajectories to identify critical traffic scenes. This can be done, for example, using classic metrics from accident research and traffic analysis (time-to-collision, etc.). If critical traffic scenes or real-world accidents are found in the dataset according to the metric used, these are excluded from step 3 and treated as "accident data." The result of this preprocessing is a normal driving trajectory dataset 401. 2) Determining uncertainties for the accidents from accident data set 405: A characteristic of data from accident databases is that the time of initial contact (primary collision) can be precisely reconstructed. In other words, the physical conditions at the time of the accident are traceable and can be represented in the trajectory(ies) of the road user(s) involved (usually as the collision point). The collision point can therefore be considered "measured" with a corresponding degree of accuracy. The temporal development of the scene before the accident (pre-crash / pre-collision phase), however, is reconstructed, and the respective (reconstructed) trajectory becomes less certain with increasing time elapsed since the accident, which is accounted for by uncertainty limits, in the example of Fig. 3 is represented as dotted lines 302, which is taken into account. To determine these uncertainty thresholds, a set of rules is first developed using expert knowledge. Based on this set of rules, uncertainty thresholds for reconstructing accidents from the accident database can be determined. The estimates can be relatively rough, but the quality of the trajectory reconstruction increases as more precise estimates become possible. If available, information from event data recorders can also be used, which determine the actual trajectory traveled in an accident, for example, from GPS coordinates. This provides an accuracy advantage over a reconstructed driving trajectory. If data from all accident participants is available from an event data recorder, a nearly exact trajectory reconstruction is possible, thus representing the highest level of accuracy. If available, information from event data recorders that provides the acceleration data of each road user immediately before the accident up to the point of collision can also be used. This can be used to estimate the accuracy limits even more precisely. 3) Model fit 402 of the normal driving trajectories from the normal driving trajectory dataset 401 according to the accident reconstruction rules. The rules from step 2 are now used to adapt the format of the recorded normal driving trajectories (normal driving trajectory dataset 401) to the reconstructed trajectories of the accident dataset (i.e., the pre-crash data) and to provide them with appropriate uncertainty limits. In the example of Fig. 3 means that the starting and ending points of a normal driving trajectory (solid curve 303 - in the example of Fig. 3 (dynamics only) are determined, and a suitable reconstructed trajectory (dashed line 301 - dynamics only) and its corresponding uncertainty limits 302 (according to the rule value from step 2) are inserted (i.e., so that the endpoints of the two trajectories 301, 303, and preferably also the starting points, coincide). This results in a processed normal driving trajectory data set 403 containing normal driving trajectories in such a way that they are comparable with the trajectories from the accident data set 405. 4) Training a machine learning model 404 (e.g., a neural network) with the processed normal driving trajectory dataset 403, which, after processing in step 3, has the same format as the accident data. Each normal driving trajectory, supplemented after step 3 with a reconstructed trajectory and the uncertainty bounds, forms a training data element. For each such training data element, the endpoint of the reconstructed trajectory (which is identical to the endpoint of the respective (recorded) normal driving trajectory) together with the uncertainty bounds (for earlier sampling times before the endpoint) is fed as input to the machine learning model 404, and the course (from beginning to end) of the normal driving trajectory is used as the target output. In other words, the machine learning model 404 learns to determine normal driving trajectories (with the same endpoint) from reconstructed trajectories (and associated uncertainty bounds).In the example of . Fig. 3 means that the behavior planner learns to determine the solid curve 303 from the dashed curve 301. For example, the use of a method like SoftMax ensures that the determined trajectory strictly adheres to the uncertainty limits. The boundary conditions are specified as interval(s), with the upper limit mapped to SoftMax output 1.0 and the lower limit to SoftMax output 0.0. This ensures that the solution generated by the network always remains within the boundary conditions. Furthermore, accident databases contain contextual information about the accident, such as the collision point represented by a GPS position or lane and road width. This contextual information (i.e., information about the scene at the time of the accident) can also be used for training. If additional information, such as obstructions to visibility or objects at the roadside, is recorded, this can also be made available to the behavior planner, thus allowing for the representation of further inherent uncertainties.Furthermore, information on weather conditions, road condition, or surface type can also be provided to the behavior planner, thus allowing for the representation of inherently greater uncertainties. The result of the training is a trained machine learning model 406. 5) Reconstruction of realistic trajectories for the accident data: In this step, the accident data set 405 (including reconstructed trajectories and uncertainty bounds describing the temporal history of the accident) is used. This data set consists of elements that each contain a reconstructed trajectory (up to an accident location or time) and uncertainty bounds, as well as, if applicable, information about the scene at the time of the accident. Each of these data elements is fed to the trained machine learning model 406, which then reconstructs the likely trajectory(ies) of the person(s) involved in the accident. Thus, with the help of the trained machine learning model 406, reconstruction artifacts (e.g., constant speed or acceleration, constant driving in the center of the lane) are removed from the accident data or at least significantly mitigated. For example, the trajectories of those involved in accidents are extrapolated in many accident databases assuming constant speed, while the trained machine learning model 406 modulates the typically human variations learned in step 4 onto the reconstruction result. This means that in the target application, i.e., when using the data set generated in this way, which is mixed with accident and normal driving data (e.g., for evaluation regarding criticality or for training another behavior planner (e.g., another neural network) for a criticality assessment), the accident trajectories can no longer be incorrectly recognized by the AI ​​simply because the vehicles are driving at a constant speed. If information from event data writers is available, it can also be used as a basis for training the 404 machine learning model, so that the network learns which uncertainties occurred during reconstruction. 6) Post-processing of scenarios: For each scenario (normal journeys and accident journeys) and the trajectories of the road users present in the scenario (for accident journeys, the trajectories determined after step 5), scene representations are calculated for the corresponding time sampling points (which are expected by further processing, e.g., an AI-based behavior planner). Furthermore, the information likely known for the respective road users at the corresponding times can be provided as a history. If static objects are also present in the data, visibility can be modeled using ray tracing methods, and objects that were invisible to a road user at the corresponding time can be removed from the scene representation or declared as invisible.In addition, based on the rule set defined in step 2, further road users can be added to the accident scenarios who may have been present in the scene but were not recorded in the respective database as uninvolved parties. For training a behavior planner (e.g., another machine learning model) to classify a scene (i.e., a traffic scene) as critical or non-critical, scenes generated in this way based on recorded trajectories are labeled as non-critical, while at least all scenarios generated in this way based on accident data are labeled as critical, thus providing appropriately labeled training examples for critical and non-critical scenes.

[0051] In summary, the training process for reconstructing trajectories from accident data in an accident database begins with the preprocessing of a dataset containing recorded trajectories. This is achieved by applying uncertainties typical of accident data to the measured trajectory points. Next, a machine learning model is trained to reconstruct the original trajectory from the endpoint and the uncertainty limits. This trained machine learning model is then used to reconstruct the trajectories of road users from the accident data. Finally, a dataset is generated containing both the reconstructed accident data and the recorded, non-critical trajectories. Scenarios reconstructed from accident data are marked as critical, while the recorded trajectories are marked as non-critical.

[0052] There are several implementation variants for the rule set from step 2. In one variant, the rule set consists of various driver-vehicle models, where a relative tolerance of, for example, 10% can be assumed for the speeds available in the dataset. It is also assumed that road users stay within their lanes unless the trajectories or accident data explicitly indicate a lane change. This results in longitudinal and lateral uncertainties for each trajectory point along the lane. Typically, the uncertainties widen over the backward extrapolated time. Alternatively, a 3-sigma limit can be used as the uncertainty boundary if the trajectory uncertainty in the data has been modeled probabilistically.

[0053] Furthermore, the rulebook may contain rules specifying the positions where other road users could have been, based on data from the accident database, and where this can potentially be ruled out. Such rules may depend on the traffic situation context, particularly the road topology. Similarly, the uncertainty thresholds may be context-dependent.

[0054] Step 4 of the processing described above essentially corresponds to a prediction task under constraints, except that in this case, time runs backward into the past rather than forward into the future. Accordingly, various prediction networks can be used here. Such a prediction task can be solved, for example, using a neural convolutional network, where the current traffic situation, including map information, is encoded as images. Other solutions are possible, for example, based on neural graph networks.

[0055] In summary, a method is provided according to various embodiments, as described in Fig. 5 shown.

[0056] Fig. Figure 5 shows a flowchart 500, which represents a method for generating training data for training an AI-based behavior planner for an at least partially automated vehicle according to one embodiment.

[0057] In 501, vehicle trajectories are provided in traffic scenes without an accident.

[0058] In section 502, accident event data elements are provided, each describing an accident event.

[0059] In 503, a machine learning model is trained on the basis of the provided vehicle trajectories without accident events to reconstruct the vehicle trajectories without accident events.

[0060] In section 504, the training data for training the behavior planner is generated by, for each accident event data element, • in 505 Generating a trajectory of at least one vehicle involved in the accident event for the accident event using the machine learning model; and • in 506 Forming a respective training data element of the training data to train the behavior planner from the generated trajectory of the at least one vehicle involved in the accident event and information that the trajectory of the vehicle involved in the accident event belongs to a critical traffic scene.

[0061] The procedure of Fig. 5. This can be performed by one or more computers with one or more data processing units. The term "data processing unit" can be understood as any type of entity that enables the processing of data or signals. The data or signals can, for example, be processed according to at least one (i.e., one or more than one) special function performed by the data processing unit. A data processing unit can include or be composed of an analog circuit, a digital circuit, a logic circuit, a microprocessor, a microcontroller, a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), an integrated circuit, a programmable gate array (FPGA), or any combination thereof.Any other method for implementing the respective functions described in more detail herein may also be understood as a data processing unit or logic circuit arrangement. One or more of the process steps described in detail herein may be executed (e.g., implemented) by a data processing unit through one or more special functions performed by the data processing unit.

[0062] According to various explanations, the process is therefore implemented primarily using computers.

[0063] The procedure according to Fig.The training data generated can be used to train a neural network that receives information about traffic scenes from various sources, such as sensor data acquired from at least one sensor. Different implementations can receive and use time series of sensor data from various sensors, such as video, radar, LiDAR, ultrasound, motion, thermal imaging, etc. Sensor data can be measured or simulated over time periods (corresponding to one or more event times and within one or more predefined events).

[0064] Based on the output of the behavior planner (which, for example, performs a criticality assessment), a device can then be controlled (e.g., driven more cautiously, such as slowing down) if the criticality determined by the behavior planner is high. While in the examples above the device is a vehicle, the approach can also be applied to other devices, generally robotic devices. For example, a robotic arm may also encounter various traffic scenarios that are more or less likely to lead to an accident (e.g., a collision with another device or a human user). The term "robotic device" can be understood as referring to any technical system (with a mechanical part whose movement is controlled and which therefore has a trajectory), such as...a computer-controlled machine, a vehicle, a household appliance, a power tool, a manufacturing machine, a personal assistant, or an access control system.

Claims

[1] Method for generating training data for training a Kl-based behavior planner for an at least partially automated vehicle, comprising: Providing (501) vehicle trajectories in traffic scenes without an accident event; Providing (502) accident event data elements each describing an accident event; Training (503) a machine learning model (404, 406) on the basis of the provided vehicle trajectories without accident events to reconstruct the vehicle trajectories without accident events; Generating (504) the training data for training the behavior planner (108) by, for each accident event data element, Generating (505) a trajectory (303) of at least one vehicle involved in the accident event for the accident event by means of the machine learning model (404, 406); and Forming (506) a respective training data element of the training data for training the behavior planner (108) from the generated trajectory (303) of the at least one vehicle involved in the accident event. [2] Method according to claim 1, wherein the respective training data element of the training data for training the behavior planner (108) is further formed from information that the trajectory of the vehicle involved in the accident event belongs to a critical traffic scene. [3] The method of claim 1 or 2, wherein the machine learning model (404, 406) is trained to reconstruct the vehicle trajectories without accident events from their endpoints and information about areas in which they extend. [4] The method of claim 3, wherein the machine learning model (404, 406) is trained to output a vehicle trajectory that lies within the range for a training input that specifies the end point of a vehicle trajectory and a range for a course of the vehicle trajectory before the end point. [5] Method according to claim 3 or 4, wherein each accident event data element contains a reconstructed trajectory (301) with associated uncertainty bounds (302) and for each training data element for the machine learning model (404, 406) the range for the course of each of the at least one vehicle trajectory is defined by uncertainty bounds (302) of a respective reconstructed trajectory (301) whose end point (300) is set to the end point of the vehicle trajectory. [6] Method according to one of claims 3 to 5, wherein for each accident event data element, the trajectory (303) of the at least one vehicle involved in the accident event is generated by determining an area for the course of the trajectory of the at least one vehicle involved in the accident event and generating the trajectory (303) of the at least one vehicle involved in the accident event by supplying the determined area and an accident location (300) as the end point of the trajectory of the at least one vehicle involved in the accident event to the machine learning model (404, 406). [7] Method according to one of claims 1 to 6, further comprising generating further training data elements for training the behavior planner (108), wherein each further training data element is formed from at least one of the vehicle trajectories in traffic scenes without an accident event or a trajectory generated by the machine learning model (108) for a traffic scene without an accident event and information that the at least one trajectory of the vehicle involved in the accident event belongs to a non-critical traffic scene. [8] Method according to one of claims 1 to 7, wherein the training data elements of the training data for training the behavior planner (108) are supplemented with position information about further objects in the respective traffic scene and / or the information for one or more further objects that they are not visible in the respective traffic scene as input for the behavior planner. [9] Method according to one of claims 1 to 8, wherein information about objects that are not visible in the respective traffic scene is removed from the training data elements of the training data for training the behavior planner (108). [10] Data processing device (105) which is arranged to carry out a method according to one of claims 1 to 9. [11] A computer program comprising instructions which, when executed by a processor, cause the processor to perform a method according to any one of claims 1 to 9. [12] A computer-readable medium storing instructions which, when executed by a processor, cause the processor to perform a method according to any one of claims 1 to 9.

Images