Computer-implemented method for training a model for analyzing a traffic scene and use of a computer-implemented model trained in this way

Abstract

A computer-implemented method for training a model for analyzing a traffic scene is proposed, wherein the model comprises a GNN encoder (31) and an analysis decoder (32) for generating analysis results (7) on the basis of the latent features generated by the GNN encoder (31). GNN (6). The training data elements for the training procedure each include a graph representation (3) and an image representation (2) of a training scene as well as an analysis result as ground truth information. For each training data element, a GNN set z is generated using the GNN encoder 31 based on the graph representation (3). GNN (6) of latent features to generate a GNN analysis result (7) with the help of the analysis decoder (32) and a distance £ analysis between the GNN analysis result (7) and the ground truth information and compare it with an optimization criterion. According to the invention, based on the image representation (2) and the GNN set z GNN (6) At least one further distance is determined from the training scene and compared with at least one further optimization criterion. The parameters of the GNN encoder (31) and / or the analysis decoder (32) are modified as long as the optimization criteria are not met.

Description

State of the art

[0001] The invention relates to a computer-implemented method for training a computer-implemented model for analyzing a traffic scene and to a computer-implemented method and system for analyzing a traffic scene with the aid of a computer-implemented model trained according to the invention.

[0002] Such a computer-implemented model comprises at least one graphical neural network referred to as a GNN encoder for generating latent features based on a graph representation of a traffic scene and an analysis decoder for generating analysis results based on the latent features generated by the GNN encoder.

[0003] For the training method, at least one set of training data elements is provided, wherein each training data element comprises at least one graph representation of a training scene and ground truth information in the form of an analysis result for this training scene.

[0004] To train the model, at least the following steps are performed for each training data element: In a first step, a GNN set of latent features is generated based on the graph representation of the training scene using the GNN encoder, i.e., a GNN representation of the training scene in a latent space. In a further step, at least one GNN analysis result is generated based on the GNN set of latent features using the analysis decoder. Then, a distance £ analysis between at least one GNN analysis result and the ground truth information and compared with an optimization criterion.

[0005] The model in question is to be trained for analyzing traffic scenes based on contextual, semantic scene understanding. Examples of applications include: - Detect whether a participant / agent in the traffic scene is parked or stopping or standing for another reason; - Identifying a possible lead vehicle for an automated follow-up journey (Automated Cruise Control); - Speed ​​adjustment, e.g. bus + children -> lower speed, no pedestrians -> higher speed; - Identify critical / dangerous areas in the traffic scene in order to avoid such areas; - Determining uncertainties of one's own system with regard to the scene in the sense of: Does my system know this scene and can my system handle this scene.

[0006] A graph neural network (GNN) is a subtype of an artificial neural network (ANN). GNNs are characterized by the fact that they process input data in a graph structure, i.e., in the form of graphs with nodes and edges, and by the fact that they apply deep learning (DL) methods. Many architectures fall into the GNN category, such as graph convolution networks (GCN), graph attention networks (GAT), and even the highly influential transformer approach can be considered a special case of a GNN – see Hamilton, “Graph Representation Learning” (2020). A GNN analyzes the input data, recognizes patterns and relationships, and can perform tasks such as classification or prediction. Typical application areas are complex problems with many dependencies and interactions between the data points, such as semantic analysis.

[0007] In the application area of ​​autonomous driving, and especially in the prediction and planning of vehicle trajectories, GNNs are distinguished by their ability to effectively model relationships and dependencies between different elements in the road environment – ​​see Gao et al., "VectorNet: Encoding HD Maps and Agent Dynamics from Vectorized Representation" (2020). This allows roads and traffic networks to be represented as graphs, with vehicles, pedestrians, and road infrastructure nodes as interconnected nodes. GNNs can leverage this graph structure to capture complex interactions between vehicles, consider the influence of neighboring entities, and make informed predictions about future trajectories. This allows autonomous vehicles to anticipate and respond to dynamic situations, such as lane merging, changes in traffic flow, or unexpected obstacles.This capability makes GNNs a powerful tool for improving the safety and efficiency of autonomous driving systems. Another key advantage of GNNs is their comparatively low computational time during inference, especially compared to convolutional neural networks (CNNs).

[0008] However, GNNs often struggle to effectively encode and understand distances or spatial relationships within a traffic scene. This is primarily because GNNs leverage the topological structure of graphs and focus on node features and their connectivity. Spatial or geometric relationships between nodes tend to be neglected. In typical graph representations, edges represent relationships or interactions without explicitly encoding distances or spatial arrangements. The lack of spatial understanding of GNNs adversely affects contextual, semantic scene understanding and thus the quality of GNN analysis results or predictions. Disclosure of the invention

[0009] The training approach according to the invention aims to improve the "spatial understanding" of a GNN-based model for analyzing a traffic scene. To this end, during training, in addition to the graph representation of a training scene, the GNN is also provided with spatial information about the respective training scene, specifically in the form of at least one image representation of the training scene. In this way, the GNN learns a spatial understanding that is then used during inference. It is crucial that no image representation of the current traffic scene is required for inference. The GNNs trained according to the invention only use a graph representation of the current traffic scene, thus retaining the advantage of GNNs' comparatively low computing time.

[0010] To achieve this, the training method according to the invention provides that each training data element further comprises at least one image representation of the training scene. Furthermore, based on the image representation and the GNN set of the training scene, at least one further distance is determined and compared with at least one further optimization criterion. As long as the optimization criterion and / or the at least one further optimization criterion are not met, training continues by modifying at least one parameter of the GNN encoder and / or the analysis decoder.

[0011] Preferably, each training scene describes a snapshot of a traffic scene or a temporal development of a traffic scene over a predetermined period of time, in particular in the form of graph representations and image representations of the traffic scene for a sequence of time steps.

[0012] In the context of the invention, an image representation of the training scene is understood to be a data representation that inherently reflects the spatial characteristics of the training scenes, in particular distances, relative orientations, and spatial relationships between participants and elements of the traffic scene. This could, for example, be a 3D voxel representation or a bird's-eye view (BEV) representation. According to the invention, the training data elements are supplemented by this type of representation of the training scenes because it allows for a very comprehensive and multifaceted spatial understanding of the training scenes and thus of traffic scenes in general.

[0013] In principle, within the scope of the invention, different further representations of this training scene can be generated on the basis of the image representation and the GNN set of a training scene and compared with each other in order to determine at least one further distance.

[0014] For this purpose, in one embodiment of the training method according to the invention, a convolutional neural network referred to as a CNN encoder is used to generate a CNN set of latent features based on the image representation of the training scene—i.e., a CNN representation of the training scene in a latent space. A vision transformer architecture, for example, could also be used as a CNN encoder, which divides the image representation of the training scene into patches for encoding. Such encoders are particularly suitable for grid-based image representations.

[0015] In this case, a first further distance in the latent space can be determined, namely a distance £ self-sup between the GNN set of latent features and the CNN set of latent features.

[0016] Based on the CNN set of latent features, at least one CNN analysis result can be generated using the analysis decoder. This allows a distance between the GNN analysis result and the CNN analysis result to be determined as a second additional distance.

[0017] Alternatively or additionally, the distance between the CNN analysis result and the ground truth information can be determined as a third additional distance.

[0018] The first, second and third additional distances are all directly or indirectly influenced by the image representation of the test scene, so that the spatial understanding of the test scene conveyed by the image representation is incorporated here.

[0019] In a further embodiment of the training method according to the invention, an image decoder is used to generate a GNN reconstruction of the image representation of the training scene based on the GNN set of latent features. In this case, a distance £ GNN-reconstruction between the image representation of the training scene and the GNN reconstruction as a fourth additional distance.

[0020] Within the scope of the invention, one could also use an image decoder to generate a CNN reconstruction of the image representation of the training scene based on the CNN set of latent features.

[0021] As a fifth further distance one could then use a distance of £ CNN-reconstruction between the image representation of the training scene and the CNN reconstruction and as a sixth further distance a distance between the GNN reconstruction and the CNN reconstruction.

[0022] The fourth, fifth and sixth additional distances are all directly or indirectly influenced by the image representation of the training scene, so that the spatial understanding of the training scene conveyed by the image representation also flows in here.

[0023] It should be noted here that different image decoders could be used for GNN reconstruction and CNN reconstruction, or even the same image decoder could be used.

[0024] Furthermore, the GNN reconstruction and / or the CNN reconstruction can be generated in the data representation of the image representation of the training scene or in another data representation that inherently reflects the spatial characteristics of the training scenes. For example, the image representation of the training scene could be a birds-eye view (BEV) representation, but a 3D voxel representation could be chosen for the GNN reconstruction and / or the CNN reconstruction.

[0025] According to the invention, the additional distances determined in the individual embodiments of the training method are each compared with a separate optimization criterion. The actual "learning" then occurs by modifying at least one parameter of the GNN encoder and / or the analysis decoder if the optimization criterion and / or the at least one additional optimization criterion is not met. Typically, parameters of the other DL components of the training architecture are also modified, in particular the CNN encoder and the image decoder(s), if present.

[0026] At this point, it should be noted that the optimization criterion and / or the further optimization criterion can simply be a minimization of the corresponding distance or, for example, a threshold value for the corresponding distance.

[0027] As already indicated at the beginning, a preferred application of a model trained according to the invention is behavior prediction and / or behavior planning for participants / agents in a traffic scene. Accordingly, the training data elements comprise at least one future behavior of at least one participant in the respective training scene as ground truth information. In this case, the analysis decoder also generates analysis results in the form of predictions or planning of the behavior of at least one participant in the respective training scene.

[0028] Using the training procedure described above, a GNN-based model for the analysis of a traffic scene can be provided that is very efficient in terms of computation time and also has a good understanding of geometric distances and relative orientations, since the GNN is trained with additional reconstruction and latent comparison tasks (self-supervision tasks).

[0029] The claimed computer-implemented method for analyzing a traffic scene, in particular for predicting and / or planning the behavior of at least one participant in a traffic scene, is characterized by the use of a model trained as described above. It comprises at least the following steps: Aggregating scene-specific information of a traffic scene and generating a graph representation of the traffic scene based on the scene-specific information,

[0030] Using the pre-trained GNN encoder to generate a set of latent features based on the graph representation of the traffic scene, using the pre-trained analysis decoder to generate an analysis result, in particular a prediction / planning of at least one behavior for at least one participant in the traffic scene, based on the set of latent features.

[0031] This procedure is used in conjunction with Fig. 3 is explained in more detail.

[0032] The claimed computer-implemented system for analyzing a traffic scene, in particular for predicting and / or planning the behavior of at least one participant in a traffic scene, is also characterized by the use of a model trained as described above. The system comprises at least the following components: a perception layer for aggregating scene-specific information of a traffic scene, an edge / node encoder for generating an initial graph representation of the traffic scene based on the aggregated scene-specific information, a model trained according to the invention with a GNN encoder for generating a set of latent features on the basis of the graph representation of the traffic scene, and with an analysis decoder for generating analysis results, in particular predictions / planning of at least one behavior for at least one participant in the traffic scene, on the basis of the set of latent features generated by the GNN encoder. The individual components of the system according to the invention are used in conjunction with the Fig. 3 and Fig. 4 is explained in more detail. Drawings

[0033] Embodiments and advantageous developments of the invention are explained in more detail below in conjunction with the figures. Fig. Figure 1 illustrates the mapping of aggregated scene-specific information of a traffic scene to a set of latent features – in the upper half of the image using a CNN encoder and in the lower half of the image using a GNN encoder. Fig. 2 shows a training architecture for carrying out a training method according to the invention using the example of a model for behavior prediction and / or planning. Fig. 3 shows the part of the Fig. 2, which is used during inference, i.e. for behavior prediction and / or planning during vehicle operation. Fig. Figure 4 shows a possible implementation of an aggregation component used in conjunction with a GNN encoder. Description of implementation examples

[0034] Fig. 1 compares a CNN encoder 10 and a GNN encoder 20 in mapping scene-specific information 1 and 2 of a traffic scene to a set of latent features.

[0035] The CNN encoder 10 receives as input an image representation of a traffic scene, which inherently reflects the spatial characteristics of the traffic scene, in particular distances and spatial relationships between participants and elements of the traffic scene. In the exemplary embodiment presented here, this is a so-called raster image from the bird's-eye view perspective (BEV image), i.e., an image representation within the meaning of the invention. In contrast, the GNN encoder 20 receives as input a graph 3, which has been generated based on scene-specific information 1 with the aid of one or more edge / node encoders 21. Such a graph 3 can be configured in different ways. However, agents in the traffic scene and / or map elements are usually represented as nodes, while their mutual relationships are modeled by edges between the corresponding nodes. Fig. 1 is the set of latent j-nodes under the representation of graph 3 z0j and i-edges e0i Features specified {z01,z02,...,e01,e02,...}.

[0036] Both encoders 10 and 20 process their respective inputs 2 and 3, respectively, and generate output representations 4 and 5 in the form of latent (encoded) features. The output representation or output of the CNN encoder 10 is usually a single latent feature vector z. CNN 4, which describes the entire traffic scene.

[0037] The GNN encoder 20 essentially performs a GNN update step, known as message passing. This involves exchanging information between the connected nodes using graph operations (Graph Attention Network / Graph Convolution Network, etc.), whereby the information content of each node changes depending on the neighboring nodes. To exchange information between more distant nodes, this update step can be repeated several times. Furthermore, the latent features of all j nodes generated by the GNN encoder 20 must be z1j and i-edges e1i with the help of an aggregation component 22 to a single latent feature vector z describing the scene GNN 6 are processed.

[0038] The aggregation component could, for example, be a multilayer perceptron (MLP) that receives all node and edge features as input. The variable number of nodes and edges is handled by filling the input vector with fill values ​​(Nans, 0, ...).

[0039] The set of latent j-nodes z1j and i-edges e1i is in Fig. 1 below the initial representation 5 {z11,z12,...,e11,e12,...}. The aggregation step is not required in the case of the CNN encoder 10 due to the missing graph structure.

[0040] Fig. Figure 1 illustrates that the scene-specific information 1 or an image representation 2 of the traffic scene cannot be used directly as input for a GNN encoder, but must first be “translated” into a graph by encoding the input data of each edge and each node individually using edge / node encoders in order to e0i,z0j to obtain, whereby e0i latent edges and z0i latent node features. Furthermore, Fig. 1, that GNN encoders, in contrast to CNN encoders, require an aggregation component to generate a latent feature vector that describes the entire traffic scene.

[0041] After these basic explanations, some embodiments of the method according to the invention for training a model for analyzing a traffic scene are described below in connection with Fig. 2. The computer-implemented GNN-based model presented here is intended to be trained to predict and / or plan the behavior of at least one participant in a traffic scene.

[0042] The model to be trained comprises a GNN encoder 31 for generating latent features in the form of a latent feature vector z GNN 6 based on a graph representation 3 of a traffic scene, as described above with reference to Fig. 1. The representation of an edge / node encoder and an aggregation component was discussed in Fig. 2 has been omitted for reasons of clarity.

[0043] Furthermore, the model to be trained comprises an analysis decoder 32 for generating analysis results based on the latent features generated by the GNN encoder 31, ie based on the latent feature vector z GNN6. In the embodiment described here, the analysis decoder 32 is a trajectory decoder that generates behavior predictions or behavior planning for individual participants in the traffic scene in the form of trajectory data 7.

[0044] For the training method, at least one set of training data elements is provided, each training data element comprising a graph representation 3 of a training scene and ground truth information in the form of at least one trajectory of at least one participant in the training scene. The ground truth information is in Fig. 2 not shown.

[0045] According to the invention, each training data element also comprises at least one image representation 2 of the training scene.

[0046] For image representation 2 of the training scenes, different data representations can be selected. However, the selected data representation should inherently reflect the spatial characteristics of the training scenes, especially distances and spatial relationships between participants and elements of the traffic scene. Suitable examples include a 3D voxel representation or a bird's-eye view (BEV) representation. Vehicles can be represented, for example, by bounding boxes. The past trajectories of agents can be represented by faded bounding boxes or using multiple image channels. Map information can be represented by different lines of varying thickness and color.

[0047] In the embodiment described here, a BEV representation was chosen for image representation 2 of the training scenes.

[0048] For each training data element, the following steps are carried out: Based on the graph representation 3 of the training scene, a GNN set of latent features in the form of a latent feature vector z is generated using the GNN encoder 31. GNN 6. Based on the latent feature vector z GNN 6, a GNN analysis result in the form of at least one GNN trajectory 7 is then generated using the trajectory decoder 32. Then, a distance £ analysis or £ trajectory between the trajectories of the GNN analysis result 7 and the ground truth information and compared with an optimization criterion. The optimization criterion can be the minimization of the distance £ trajectory or can be specified by a threshold value for the distance.

[0049] According to the invention, based on the image representation 2 of the training scene and the GNN set of the training scene - latent feature vector z GNN6 - at least one further distance is determined and compared with at least one further optimization criterion. The parameters or at least one parameter of the GNN encoder 31 and / or the analysis decoder 32 are then modified during the training process as long as the optimization criterion and / or the at least one further optimization criterion are not met.

[0050] To determine further distances based on the image representation 2 and the GNN set 6 of the training scene, the Fig. 2 further comprises a CNN encoder 33 which, based on the image representation 2, generates a CNN set of latent features in the form of a latent feature vector z CNN 4 generated.

[0051] This opens up the possibility of setting a distance of £ as the first further distance. self-sup between the GNN set of latent features - feature vector z GNN6 - and the CNN set of latent features - feature vector z CNN 4 - to be determined £ self-sup = ||z CNN - z GNN ||.

[0052] In addition, the trajectory decoder 32 can, based on the feature vector z CNN 4 generate at least one CNN analysis result. A second additional distance can then be determined, namely a distance between the trajectories of the GNN analysis result and the trajectories of the CNN analysis result. However, a third additional distance can also be determined, namely a distance between the trajectories of the CNN analysis result and the ground truth information.

[0053] Furthermore, the Fig. 2, an image decoder 34, which is based on the GNN set of latent features - feature vector z GNN6 - generates a GNN reconstruction 8 of an image representation of the training scene. The data representation of image representation 2 of the training data element can be selected—here, a BEV representation—or another data representation that inherently reflects the spatial characteristics of the training scenes. For example, the image decoder 34 could also generate reconstructions in the form of semantic maps of the training scene.

[0054] As a fourth additional distance a distance of £ GNN-reconstruction between the image representation 2 of the training scene and the GNN reconstruction 8.

[0055] By reconstructing BEV images or other image representations of the test scene from the aggregated latent GNN features z GNN6, the GNN-based model can better learn the geometric relationships (distances and orientations) during training because these relationships are present in the BEV images.

[0056] The image decoder 34 or another image decoder not shown here could also be based on the CNN set of latent features - feature vector z CNN 4 - generate a CNN reconstruction of the image representation of the training scene. In this case, the fifth additional distance could be a distance £ CNN-reconstruction between the image representation 2 of the training scene and the CNN reconstruction.

[0057] Since the parameters of all participating DL components of the training architecture are usually modified during the training process, the fourth and fifth distances contribute to training the image decoder(s) 34 in such a way that the reconstructed image representations correspond as closely as possible to the image representation 2 of the training data set, even if these image representations are present in different data representations.

[0058] Finally, in the present embodiment, a sixth further distance could also be determined, namely a distance between the GNN reconstruction and the CNN reconstruction.

[0059] In summary, we can say: Due to the additional distances and optimization criteria, the GNN-based model with GNN encoder 31 and analysis or trajectory decoder 32 receives an additional signal during training, through which it implicitly learns a geometric understanding similar to that of CNN-based models.

[0060] As already mentioned, Fig. 3 the part of Fig. 2, which is used for inference, i.e. for analyzing a traffic scene during vehicle operation, or in the present embodiment for behavior prediction and / or behavior planning.

[0061] A computer-implemented system according to the invention for performing such an analysis or behavior prediction / planning comprises a perception layer for aggregating scene-specific information from the traffic scene, usually from different information sources. These can be on-board sensors, such as lidar sensors, radar sensors, and / or RGB cameras installed on the ego vehicle, or external sensors, such as lidar sensors, radar sensors, and / or RGB cameras installed in or on infrastructure elements. Other possible information sources include stored map information as well as queryable weather and road condition information, traffic situation information, etc. The information from the different information sources is aggregated by the perception layer and preprocessed into context information.

[0062] Furthermore, the system according to the invention comprises an edge / node encoder for generating an initial graph representation 3 of the traffic scene on the basis of the aggregated scene-specific information, as well as the Fig. 3 illustrated components trained according to the invention, namely the GNN encoder 31 for generating a set of latent features z GNN 6 on the basis of the graph representation 3 of the traffic scene, and the analysis or trajectory decoder 32 for generating trajectories 7 as predictions / plannings of at least one behavior for at least one participant in the traffic scene, on the basis of the set of latent features z generated by the GNN encoder 31. GNN 6.

[0063] Since only the GNN components 31 and 32 trained according to the invention are used during inference, i.e. during runtime in the vehicle, both the computational effort and the computational time are low compared to CNN-based models.

[0064] In Fig. 4 is a possible embodiment of the aggregation component 22 from Fig. 1 shown.

[0065] The aggregation component 22 is implemented here in the form of another GNN-based network, in which a middle, uninitialized node is connected in a star-shaped manner to all other nodes of the graph, while no further connections exist between the nodes of the previous graph. This form of aggregation component is described, for example, in Janjos et al., "SAN: Scene Anchor Networks for Joint Action-Space Prediction" (2022). After the GNN update step (message passing, ...), the latent features of the middle node can then be aggregated as latent features. GNN be considered. QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions. Cited non-patent literature

[0000] Hamilton, “Graph Representation Learning” (2020

[0006] Gao et al., “VectorNet: Encoding HD Maps and Agent Dynamics from Vectorized Representation” (2020

[0007] Janjos et al. “SAN: Scene Anchor Networks for Joint Action-Space Prediction” (2022

[0065]

Claims

[1] A computer-implemented method for training a model for analyzing a traffic scene, the model comprising at least: • A graphical neural network called GNN encoder (31) for generating latent features z GNN (6) based on a graph representation (3) of a traffic scene and • an analysis decoder (32) for generating analysis results (7) on the basis of the latent features generated by the GNN encoder (31) GNN (6); wherein at least one set of training data elements is provided for the training method, each training data element comprising at least: • a graph representation (3) of a training scene and • ground truth information in the form of an analysis result for this training scene; and wherein for each training data item, at least the following steps are carried out: a) Using the GNN encoder (31) to generate a GNN set z GNN (6) of latent features based on the graph representation (3) of the training scene, b) using the analysis decoder (32) to generate at least one GNN analysis result (7) based on the GNN set z GNN (6) of latent characteristics, and c) Determine a distance £ analysis between the at least one GNN analysis result (7) and the ground truth information and comparing the distance £ analysis with an optimization criterion; characterized by , that each training data element further comprises at least one image representation (2) of the training scene, that based on the image representation (2) and the GNN set z GNN (6) at least one further distance is determined from the training scene and compared with at least one further optimization criterion and that at least one parameter of the GNN encoder (31) and / or the analysis decoder (32) is modified if the optimization criterion and / or the at least one further optimization criterion is not met. [2] Method according to claim 1, characterized by that based on the image representation (2) of the training scene, a CNN set z is generated with the help of a convolutional neural network called a CNN encoder (33). CNN (4) is generated from latent features. [3] Method according to claim 2, characterized by that the first further distance is a distance of £ self-sup between the GNN set z GNN (6) of latent features and the CNN set z CNN (4) is determined by latent characteristics. [4] Method according to one of claims 2 or 3, characterized by that based on the CNN set z CNN(4) at least one CNN analysis result is generated from latent features with the aid of the analysis decoder (32), and that a distance between the GNN analysis result and the CNN analysis result is determined as a second further distance and / or that a distance between the CNN analysis result and the ground truth information is determined as a third further distance. [5] Method according to one of claims 1 to 4, characterized by that based on the GNN set z GNN (6) of latent features with the help of an image decoder (34) a GNN reconstruction (8) of the image representation (2) of the training scene is generated and that as a fourth further distance a distance £ GNN-reconstruction between the image representation (2) of the training scene and the GNN reconstruction (8). [6] Method according to one of claims 2 to 5, characterized by that based on the CNN set z CNN(4) a CNN reconstruction of the image representation of the training scene is generated from latent features using an image decoder and that as a fifth further distance a distance of £ CNN-reconstruction between the image representation (2) of the training scene and the CNN reconstruction is determined and / or that as a sixth further distance a distance between the GNN reconstruction (8) and the CNN reconstruction is determined. [7] Method according to one of claims 1 to 6, characterized by that each training scene describes a snapshot of a traffic scene or a temporal development of a traffic scene over a predetermined period of time, in particular in the form of graph representations and image representations of the traffic scene for a sequence of time steps. [8] Method according to one of claims 1 to 7, characterized bythat a data representation is chosen as the image representation of the training scenes that inherently reflects the spatial characteristics of the training scenes, in particular distances and spatial relationships between participants and elements of the traffic scene, in particular in the form of a 3D voxel representation or a Birds-Eye-View (BEV) representation. [9] Method according to one of claims 5 or 6 and 8, characterized by that the GNN reconstruction and / or the CNN reconstruction is generated in the data representation of the image representation of the training scene or in another data representation that inherently reflects the spatial characteristics of the training scenes. [10] Computer-implemented method according to one of claims 1 to 9, with which a model for predicting and / or planning a behavior of at least one participant in a traffic scene is trained, wherein the training data elements comprise at least one future behavior of at least one participant of the respective training scene as ground truth information, and whereby, with the help of the analysis decoder, analysis results are generated in the form of predictions or planning of the behavior of at least one participant in the respective training scene. [11] Computer-implemented method for analyzing a traffic scene, in particular for predicting and / or planning a behavior of at least one participant in a traffic scene, with the aid of a model trained according to one of claims 1 to 10, comprising the following steps: a. Aggregating scene-specific information of a traffic scene and generating a graph representation (3) of the traffic scene based on the scene-specific information, b. Using the pre-trained GNN encoder (31) to generate a set of latent features z GNN(6) based on the graph representation (3) of the traffic scene, c. Using the pre-trained analysis decoder (32) to generate an analysis result (7), in particular a prediction / planning of at least one behavior for at least one participant in the traffic scene, based on the set of latent features z GNN (6). [12] Computer-implemented system for analyzing a traffic scene, in particular for predicting and / or planning the behavior of at least one participant in a traffic scene, the system comprising at least the following components: a. a perception layer for aggregating scene-specific information of a traffic scene, b. an edge / node encoder (21) for generating an initial graph representation (3) of the traffic scene based on the aggregated scene-specific information, c. a model trained according to any one of claims 1 to 10 i. with a GNN encoder (31) for generating a set z GNN (6) of latent features based on the graph representation (3) of the traffic scene, and ii. with an analysis decoder (32) for generating analysis results (7), in particular predictions / planning of at least one behavior for at least one participant in the traffic scene, on the basis of the set z generated by the GNN encoder (31). GNN (6) of latent traits. [13] Vehicle with a computer-implemented system for analyzing a traffic scene, in particular for predicting and / or planning a behavior of at least one participant in a traffic scene, according to claim 12.

Images