Computer-implemented method and system for behavioral planning of traffic scene participants

By distributing scene-specific information over multiple semantic levels and using a DL architecture to map these relationships, the method addresses the challenge of modeling distant semantic relationships in traffic scenes, improving prediction and behavior planning accuracy.

CN119068451APending Publication Date: 2024-12-03ROBERT BOSCH GMBH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410687985.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2023-05-31
Filing Date
2024-05-30
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

Existing deep learning-based methods for autonomous driving struggle to effectively model the semantic relationships between distant areas of a traffic scene, leading to diluted and incomplete information for prediction and behavior planning.

Method used

Distribute scene-specific information over multiple semantic levels of a grid-based scene representation and divide blocks accordingly, using a DL architecture to map these relationships onto latent features, considering the specific semantics of sub-blocks and their interactions.

Benefits of technology

Enhances the extraction and utilization of relevant information for prediction and planning tasks by maintaining semantic distinctions, allowing for more accurate and comprehensive behavior planning in dynamic traffic scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119068451A_ABST
    Figure CN119068451A_ABST
Patent Text Reader

Abstract

A computer-implemented method and system for behavioral planning of traffic scene participants is presented. To this end, a mesh-based scene representation (30) is generated on the basis of the aggregated scene-specific information and is divided into a plurality of blocks (35), each representing a sub-region of the traffic scene. According to the invention, the scene-specific information is distributed over at least two semantic levels of the scene representation (30). The distribution is maintained when divided into a plurality of chunks (35) to form at least one type of sub-chunks (351, 352, 353, 354) based on these chunks (35) and semantic levels. Then, the mesh-based scene representation (30) is mapped to potential features based on these tiles (35) or sub-tiles (351, 352, 353, 354), where semantic relationships between at least one type of tiles (35) or sub-tiles (351, 352, 353, 354) are taken into account. Potential features generated in this way are used as inputs for at least one downstream DL (deep learning) module to predict the development of a traffic scene and / or for behavioral planning.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a computer-implemented method for the behavior planning of at least one participant in a traffic scene, wherein firstly at least one grid-based scene representation is generated based on aggregated scene-specific information The grid-based scene representation is divided into a number of tiles, which represent sub-areas of the traffic scene, respectively. Based on these tiles, the grid-based scene representation is then mapped onto latent features, taking into account the semantic relations between the tiles. The latent features generated in this way are used as input to at least one downstream deep learning (DL) module for predicting the development of the traffic scene and / or for behavior planning.

[0002] The present invention also relates to a computer-implemented system for behavior planning of at least one participant of a traffic scene. The system comprises: a perception module for generating a grid-based scene representation based on aggregated scene-specific information; and a preprocessing component for dividing the grid-based scene representation into a plurality of blocks, each of which represents a sub-area of ​​the traffic scene. In addition, the system comprises a DL architecture for mapping the grid-based scene representation to latent features by using the blocks and taking into account semantic relations between the individual blocks of the grid-based scene representation. The system also comprises at least one downstream DL module for performing behavior planning and / or predicting the development of the traffic scene based on the latent features. Background Art

[0003] To achieve high-performance and safe autonomous driving, autonomous vehicles must be able to anticipate future developments in driving situations and take corresponding actions. Classic model-based approaches have reached their limits, especially in various urban scenarios, making DL-based methods for prediction and behavior planning increasingly mature.

[0004] The method discussed here is based on a grid-based representation of the current traffic scene. To this end, scene-specific information is aggregated from different information sources of the ego vehicle and, if necessary, also from information sources outside the vehicle. Typically, scene-specific information is data detected by camera sensors, lidar sensors and / or radar sensors. In addition to this, specific scene information can also be detected with the help of inertial sensors of traffic participants. Scene-specific information is usually also supplemented by GPS data and environmental information, such as weather data and road condition data. Scene-specific information is aggregated at a given point in time and therefore includes current sensor data and other data at the current point in time. However, they can also include sensor data and other data collected during a predetermined time period until a given point in time. Scene-specific information can be converted into a grid tensor (Grid-Tensor) by using map information in the form of, for example, a high-precision map, which then acts as a grid-based scene representation around the ego vehicle. Such a grid tensor can include raw sensor data, such as lidar and / or radar sensor data, or can also include sensor data in a preprocessed form, such as in the form of trajectories of corresponding participants in the traffic scene. Each mesh of this grid-based scene representation corresponds to a sub-area of ​​the traffic scene. They are referred to below as sub-blocks or "patches". It should be noted here that the individual blocks of the scene representation can represent sub-areas of different sizes and / or shapes. Therefore, the blocks do not necessarily have to be arranged in a uniform grid. It is also conceivable that there is a certain degree of overlap between the individual blocks of the scene representation. In this case, for example, position coding can be omitted, since it can be directly derived from each block which blocks it is adjacent to. In addition, objects or structures that are located exactly at the block boundaries can possibly be better recognized / taken into account.

[0005] Some examples of grid-based scene representations known in practice are stated below:

[0006] - RGB grid, where different color values ​​are used to show different semantic information, such as lane markings, vehicles, pedestrians, etc.,

[0007] -Binary The grid contains a binary layer in the grid tensor for each semantic information, where the value is 1 where the information exists and 0 otherwise.

[0008] - a voxel grid where the lidar point cloud is shown, and

[0009] - Voxel flow grid showing radar reflections.

[0010] It is also known to further process the grid tensor with the aid of a so-called backbone network to generate potential features or characteristics (Features) as input for downstream DL planning and prediction modules. Convolutional neural networks (CNNs) known from image processing have been established for this purpose.

[0011] However, a new paradigm has been developed in image processing - the so-called Vision Transformer. Vision Transformers represent an extension of the classical Transformers known from language processing (Natural Language Processing). Such Vision Transformers are described, for example, in Dosovitskiy et al., “An Image is Worth 16x16Words: Transformers for Image Recognition at Scale”. https: / / arxiv.org / abs / 2010.11929 and “Attention is all you need” in Vaswani et al. https: / / arxiv.org / abs / 2010.11929 Described in .

[0012] A major disadvantage of CNN is that the individual blocks are processed independently of the corresponding image content. The individual CNN filters slide evenly over the image, so that the relationship between adjacent pixels is primarily considered. The "perceptual field" of CNN is more local here. Although the relationship between distant areas of the image can be considered to a certain extent by subsampling between different layers of CNN. However, this information is "softened" through multiple layers of CNN.

[0013] The visual transformer addresses this shortcoming of CNNs with the help of the transformer's attention. In the case of the visual transformer, the "perception field" can therefore be globally extended across the entire image, since the attention to all patches is taken into account. Attention takes into account the locality of perception by modeling the semantic relationships between the image contents of the individual patches. To this end, the individual patches are embedded with the help of a neural network. These embeddings are then fed to the transformer. In the transformer, these patches or embeddings are interpreted as so-called tokens in order to model the relationships between the patches in the form of attention. In this way, for example, information from the upper left corner of the image can be perfectly combined with information from the lower right corner of the image and placed in a meaningful context for the respective image processing task. Summary of the invention

[0014] The present invention exploits this aspect of visual transformers in order to map information that is relevant for prediction and behavior planning in a given traffic scene and that is distributed over multiple blocks of a grid-based scene representation onto latent features in a targeted manner. This relevant information is thus made available to downstream DL modules, where it can be used for prediction and planning tasks. Known visual transformers always only consider the information content of the individual blocks as a whole in order to model the relationships between the individual blocks. According to the present invention, it has been recognized that the relevant information of the traffic scene is often "diluted" in this way. ", and therefore may not be fully mapped to the latent features. Therefore, the present invention proposes to use the information content of each block in a semantically differentiated manner. In this way, different semantic relationships between the image contents of each block can also be extracted and mapped to the latent features.

[0015] According to the invention, this is achieved by distributing the scene-specific information over at least two semantic levels of the scene representation and maintaining this distribution during the division into a plurality of blocks. Based on the blocks and the semantic levels, at least one type of sub-blocks is formed, wherein the sub-blocks of one type are respectively formed by defining subsets of the type of at least two semantic levels of the block. Semantic relationships between the sub-blocks of at least one type are then taken into account when mapping the grid-based scene representation onto the latent features.

[0016] Therefore, the core idea of ​​the present invention is to distribute the aggregated scene-specific information to different semantic levels of the grid-based scene representation and to divide each block according to the semantics of these layers. Depending on the type, the sub-blocks defined in this way can include only one semantic level or also multiple semantic levels. According to the present invention, the relationships within one or more semantic levels can be modeled in a targeted manner, while ignoring information from other semantic levels. In this way, relevant information for the prediction task and / or planning task can be extracted from the scene representation in a targeted manner. In addition, the measures according to the present invention help to make these "relevant" information also reflected in the potential features. For example, for predictive driving, the semantic relationships between the blocks at the level of infrastructure and road topology, and more precisely at the edge of the field of vision, should be particularly considered. On the contrary, for reactive driving, more attention should be paid to other traffic participants, especially the driver in front. Therefore, the semantic relationships between the sub-blocks at the infrastructure level are usually different from the semantic relationships at the level of traffic participants, but in essence they should be reflected in the potential features of the scene representation.

[0017] It should be noted here that when mapping the grid-based scene representation onto latent features, not only semantic relations between sub-patches of the same type can be taken into account, but also between sub-patches of different types. This can prove advantageous, for example, for dynamic objects. In this case, a first type of sub-patches can be used to model interactions between objects, while a second type of sub-patches models the infrastructure and thus provides constraints on the movement possibilities of the objects, such as "the vehicle must stay on the roadway".

[0018] It is important for applying the measures according to the invention that a grid-based scene representation is generated based on the aggregated scene-specific information. This may include raw sensor data, such as lidar and / or radar data. Alternatively or in addition to this, the grid-based scene representation may also include the results of a semantic evaluation of the scene-specific raw sensor data, such as object information in combination with position data, orientation data and motion data, map information, etc., of a perception module.

[0019] It is also important for the present invention that the aggregated scene-specific information is distributed over at least two different semantic levels of the scene representation. In view of the informed (avisiert) prediction and planning tasks, it has proven to be advantageous if at least one first semantic level of the scene representation includes position and / or movement information from one or more participants in the traffic scene. At least one second semantic level of the scene representation can then include information about static objects and / or infrastructure elements in the traffic scene. Within the scope of the present invention, the scene-specific information can also be divided up in other ways. For example, a separate semantic level can be set for each participant or for all participants of a type, i.e. four-wheeled vehicles, two-wheeled vehicles, pedestrians, or only for the ego vehicle, while the information of the remaining participants in the traffic scene is aggregated (zusammenfassen) at a common semantic level.

[0020] In a preferred embodiment of the present invention, the grid-based scene representation is shown in the form of a three-dimensional scene tensor, wherein the first two dimensions of the scene tensor represent the spatial area of ​​the traffic scene, and the various layers of the scene tensor in the third dimension are formed by at least two semantic levels of the scene representation. In this case, the division into multiple blocks and multiple sub-blocks is designed to be particularly easy. In order to be divided into multiple blocks, it is only necessary to divide the scene tensor into the first dimension and / or the second dimension while maintaining the third dimension of the scene tensor. The sub-scene tensors thus generated represent sub-areas of the traffic scene respectively and include all semantic levels of the scene tensor. Then, multiple sub-blocks are simply formed from one or more semantic levels of these sub-scene tensors.

[0021] An advantageous extension of the method according to the invention also takes into account causal relationships in the temporal development of traffic scenes by taking into account at least one further grid-based scene representation of at least one earlier point in time when generating latent features for these grid-based scene representations, wherein in particular the temporal sequence of these grid-based scene representations is taken into account. In order to ensure causal relationships, a so-called "attention mask" can be introduced here. For example, a mask can be used that allows blocks from time points t and t-1 to notice blocks from time point t-2, but not vice versa.

[0022] In order to map the grid-based scene representation onto latent features, a DL architecture is preferably used within the scope of the method according to the invention, which in a first step generates embeddings for the sub-blocks, usually using a positional encoding, which reproduces the position of the corresponding sub-block within the scene representation. In a second step, the DL architecture then generates latent features based on the embedding generated in this way. This may be an attention-based DL architecture, such as a visual transformer, which determines the attention with the help of a key vector, a query vector or a value vector. However, it is also conceivable to use a DL architecture that uses another predetermined measure for the relationship between the blocks and sub-blocks of the scene representation, for example as a sum or an element-by-element product.

[0023] Furthermore, a computer-implemented system for behavior planning of at least one participant of a traffic scene is claimed, which is designed to implement the measures according to the invention. Such a system comprises: a perception module for generating a grid-based scene representation based on aggregated scene-specific information; a preprocessing component for dividing the grid-based scene representation into a plurality of tiles, the tiles representing sub-areas of the traffic scene respectively; a DL architecture for mapping the grid-based scene representation onto latent features by using the tiles and taking into account semantic relations between the tiles of the grid-based scene representation; and at least one downstream DL module for behavior planning and / or predicting the development of the traffic scene based on the latent features.

[0024] According to the present invention, the perception module is designed to distribute scene-specific information to at least two semantic levels of the grid-based scene representation. In addition, the preprocessing component is designed to maintain the distribution of scene-specific information to at least two semantic levels when divided into multiple blocks and form at least one type of sub-blocks, wherein a type of sub-blocks respectively includes: a subset of the at least two semantic levels of the block defining the type. Finally, the DL architecture is also designed to consider the semantic relationship between at least one type of sub-blocks when mapping the grid-based scene representation to the latent features.

[0025] In a preferred embodiment of the present invention, in order to map the grid-based scene representation to the latent features, the DL architecture comprises: at least one neural embedding network for generating an embedding for one type of sub-block and at least one processing network for generating latent features based on the embedding so generated. It is particularly advantageous that the DL architecture comprises multiple neural embedding networks for generating embeddings for different types of sub-blocks. In this case, the processing network should be designed to take into account the embeddings of different types of sub-blocks when generating the latent features. In this way, different semantic relationships between these blocks and sub-blocks of the scene representation can be specifically identified / extracted and taken into account when mapping to the latent features.

[0026] If the processing network is designed to take into account at least one further grid-based scene representation at at least one earlier point in time when generating potential features for a grid-based scene representation, and in particular a time sequence of these grid-based scene representations, certain causal relationships in the temporal development of the traffic scene can also be taken into account. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Measures and preferred embodiments according to the invention are explained in more detail below with reference to the accompanying drawings.

[0028] Figure 1 The method of operation of the computer-implemented system according to the invention for behavior planning of participants in a traffic scenario is explained using a block diagram.

[0029] Figure 2 We illustrate our approach to generating latent features from a grid-based scene representation by using a visual transformer.

[0030] Figure 3 Use and Figure 1 The same type of representation is used to describe the measures according to the invention and in particular to show the Figure 2 The differences in methods are shown. DETAILED DESCRIPTION

[0031] Figure 1The block diagram shows a computer-implemented system 100 for behavior planning of an at least partially automated vehicle, also referred to below as an ego vehicle. The system 100 comprises a perception module 1 for generating a grid-based representation of a traffic scene. For this purpose, the perception module 1 aggregates scene-specific information 10 from different vehicle-own information sources and possibly also from vehicle-external information sources. The scene-specific information 10 is typically data detected by camera sensors, lidar sensors and / or radar sensors. In addition to this, the scene-specific information 10 can also be detected with the aid of inertial sensors of the ego vehicle. These scene-specific information 10 are typically also supplemented by GPS data and map information, weather data and road condition data, which can be provided, for example, via the cloud. These scene-specific information 10 aggregated at a given point in time typically include current sensor data and other data at that point in time. However, they can also include sensor data and other data collected over a predetermined time period up to a given point in time.

[0032] The perception module 1 may evaluate the aggregated scene-specific information 10, for example to identify and locate objects and participants in the current traffic scene. Within the scope of the evaluation, the aggregated scene-specific information 10 is converted into a grid-based scene representation by using map information, such as in e.g. Figure 2 and Figure 3 The scene tensor 20 or 30 shown in FIG. 1 is a scene-specific information 10 distributed here on at least two semantic levels of a grid-based scene representation.

[0033] Furthermore, the system 100 according to the invention comprises a pre-processing component 2 for dividing the grid-based scene representation into a plurality of blocks, which respectively represent sub-areas of the traffic scene. When dividing into a plurality of blocks, the distribution of the scene-specific information on at least two semantic levels is maintained. Another function of the pre-processing component 2 is to form at least one type of sub-blocks 11, wherein a type of sub-block 11 respectively comprises a subset of at least two semantic levels of the block defining the type. Thus, a first type of sub-block comprises only one specific semantic level, while a second type of sub-block is formed by only one other semantic level. Another type of sub-block may comprise multiple semantic levels.

[0034] These functions of the pre-processing component 2 may be fully or only partially implemented in the perception module 1. However, the pre-processing component 2 may also be an independent component of the system 100.

[0035] The system 100 further comprises a DL architecture 3 for mapping the grid-based scene representation onto latent features 12 of at least one downstream DL module 6 for predicting the development of the traffic scene and / or for behavior planning. The sub-blocks 11 are used as input data for the DL architecture 3. The DL architecture is configured and trained in such a way that semantic relations between the individual blocks or between at least one type of sub-blocks are reflected in the generated latent features 12. For this purpose, the DL architecture 3 comprises, for each type of sub-block 11, at least one neural embedding network 4 for generating an embedding for the respective type of sub-block 11 and at least one processing network 5 for generating latent features 12 based on the embedding generated in this way.

[0036] Figure 2 In the left half of FIG. 1 , a grid-based scene representation of a traffic scene is shown in the form of a three-dimensional scene tensor 20, which is combined with Figure 1 As described, it is generated by the perception module based on the scene-specific information. When converting the aggregated scene-specific information into a three-dimensional scene tensor 20, the map information is used so that the first two dimensions x and y of the scene tensor 20 represent the spatial area of ​​the traffic scene. In addition, this scene-specific information is distributed to multiple semantic levels, in the current case to a total of ten semantic levels. These semantic levels form the individual layers of the scene tensor 20 in the third dimension z. The ten semantic levels are summarized in three groups 21, 22 and 23. The first group 21 includes three semantic levels, which have information about road topology, infrastructure and navigation, the second group 22 includes four semantic levels, which have information about other participants in the traffic scene, and the third group 23 includes three semantic levels, which have information about the self-vehicle.

[0037] In already Figure 2 In the right half of Figure 2 In a further processing step performed in the left half of the diagram, the grid-based scene representation is divided into a plurality of blocks, each of which represents a sub-area of ​​a traffic scene. Here, the distribution of the scene-specific information to the different semantic levels of the scene representation is maintained. In the embodiment shown here, the scene tensor 20 is divided into nine uniform blocks or sub-scene tensors 25 in the following way: the first dimension x and the second dimension y of the scene tensor 20 are divided into three identical parts, while the individual levels in the third dimension z of the scene tensor 20 are retained. Therefore, each sub-scene tensor 25 represents a sub-area of ​​the traffic scene having all the semantic levels of the scene tensor 20.

[0038] Figure 2The method shown in provides that an embedding is generated for each block, that is, for each of the nine sub-scene tensors 25, and more precisely, the embedding is generated with the aid of a correspondingly trained embedding network 27. For this purpose, all semantic levels of the sub-scene tensor 25 are respectively supplied to the embedding network 27. Thus, an embedding for a sub-scene tensor 25 is generated based on the totality of its scene-specific information, which is here all treated equally. All nine embeddings generated in this way form input data for a processing network 28, which then generates potential features for a downstream DL-based prediction and / or planning module based on the totality of these embeddings.

[0039] Combined with the above Figure 2 Unlike the method described above, when generating latent features within the scope of the method according to the invention, these scene-specific information of the blocks are not all treated equally, but are considered in a semantically distinct way. This will be discussed below in conjunction with Figure 3 Explain in more detail.

[0040] and Figure 2 similar, Figure 3 The right half of shows the following scene tensor 30, which also comprises a total of ten semantic levels. Here it is indicated: a division into nine uniform blocks or sub-scene tensors 35. The individual blocks 35 are additionally divided into sub-blocks 351, 352, 353, 354. According to the invention, this division is performed from a semantic point of view. Therefore, the different layers of the binary scene tensor are grouped according to their semantics. If the scene tensor is an RGB grid tensor, then it is suitable to: divide by color. In addition, if earlier scene representations are also taken into account when processing the scene representation at a given point in time, a grouping by time can also be performed. In Figure 3 The division into sub-blocks 351, 352, 353, 354 is shown in the left half of . Sub-block 351 includes multiple semantic levels, which mainly contain information about static components of the traffic scene, such as infrastructure and road topology. In contrast, the semantic levels with information about participants or objects in the traffic scene are also grouped according to their time points. Three types of sub-blocks with information about the own vehicle and other traffic participants are provided here: sub-block 352 at the current time point, sub-block 353 at the previous time point, and sub-block 354 at the previous time point.

[0041] It is important for the invention that at least one type of sub-block is formed based on the block and the semantic level and possibly the time point. The sub-blocks are thus formed from subsets of the semantic level of the scene representation, wherein the respective subset determines the type of the sub-block.

[0042] According to the invention, embeddings are generated for each sub-block 351, 352, 353 and 354, respectively, and to be precise are generated with the aid of a correspondingly trained embedding network 37. Since each of these different types of sub-blocks uses its own, correspondingly trained embedding network 37, the scene-specific information of the block is already semantically differentiated during the embedding. In the present embodiment, four times nine sub-blocks are considered and four times nine embeddings are generated accordingly. These form the input data for the processing network 38, which then generates latent features for the downstream DL-based prediction and / or planning modules based on the totality of these embeddings. In this way, when mapping the grid-based scene representation onto the latent features, semantic relationships between at least one type of sub-block are taken into account. For example, a very refined attention to different segments of the road can thus be achieved. Since the embedding of different types of sub-blocks is performed with differently trained embedding networks, the method according to the invention is similar to Figure 2 Compared with the method shown in , it can fully exploit the semantics of the scene representation on a larger scale. Figure 2 The method shown in treats all scene-specific information of each block equally and does not perform semantic differentiation.

Claims

1. A computer-implemented method for behavior planning of at least one participant in a traffic scenario, a. Among them, generating at least one grid-based scene representation (30) based on the aggregated scene-specific information, b. Among them, The grid-based scene representation (30) is divided into a plurality of blocks (35), each of which represents a sub-area of ​​the traffic scene. c. wherein the grid-based scene representation (30) is mapped onto latent features based on the blocks (35), wherein semantic relationships between the blocks (35) are taken into account, and d. wherein the latent features generated in this way are used as input to at least one downstream DL (deep learning) module for predicting the development of the traffic scenario and / or for behavior planning; It is characterized in that distributing the scene-specific information onto at least two semantic levels of the scene representation and maintaining the distribution when dividing into a plurality of partitions (35), At least one type of sub-blocks (351, 352, 353, 354) is formed based on the blocks (35) and the semantic levels, wherein one type of sub-blocks (351, 352, 353, 354) is formed by defining a subset of the type at at least two semantic levels of the blocks (35), respectively, and when mapping the grid-based scene representation (30) to the latent features, a semantic relationship between at least one type of sub-blocks (351, 352, 353, 354) is considered.

2. The method according to claim 1, It is characterized in that When mapping the grid-based scene representation (30) onto the latent features, semantic relationships between sub-blocks (351, 352, 353, 354) of the same type and / or between sub-blocks (351, 352, 353, 354) of different types are considered.

3. The method according to claim 1 , It is characterized in that The grid-based scene representation comprises scene-specific raw sensor data and / or results of a semantic evaluation of the scene-specific raw sensor data.

4. The method according to any one of claims 1 to 3, It is characterized in that At least one first semantic level of the scene representation comprises position and / or motion information from one or more participants of the traffic scene, and at least one second semantic level of the scene representation comprises information about static objects and / or infrastructure elements in the traffic scene.

5. The method according to any one of claims 1 to 4, It is characterized in that The grid-based scene representation (30) is shown in the form of a three-dimensional scene tensor (30), wherein the first two dimensions (x, y) of the scene tensor (30) represent the spatial area of ​​the traffic scene, and the individual layers of the scene tensor (30) in the third dimension (z) are formed by at least two semantic levels of the scene representation (30).

6. The method according to claim 5, It is characterized in that The scene representation (30) is divided into a plurality of blocks (35) in the following manner: the scene representation is divided into a first dimension (x) and / or a second dimension (y) of the scene tensor (30), while maintaining a third dimension (z) of the scene tensor (30), so that the generated sub-scene tensors (35) respectively represent sub-areas of the traffic scene having all semantic levels of the scene tensor (30).

7. The method according to any one of claims 1 to 6, It is characterized in that When generating the potential features for the grid-based scene representation, at least one further grid-based scene representation of at least one earlier point in time is taken into account, wherein in particular a time sequence of the grid-based scene representations is taken into account.

8. The method according to any one of claims 1 to 7, It is characterized in that In order to map the grid-based scene representation (30) onto latent features, a DL architecture (37, 38) is used which generates in a first step an embedding for the sub-blocks and, in a second step, generates the latent features based on the embedding generated in this way.

9. A computer-implemented system (100) for behavior planning of at least one participant in a traffic scenario, the system comprising at least include: a. a perception module (1) for generating a grid-based scene representation based on aggregated scene-specific information (10), b. a preprocessing component (2), for dividing the grid-based scene representation into a plurality of blocks, each of which represents a sub-area of ​​the traffic scene; c. a DL (deep learning) architecture (3) for mapping the grid-based scene representation to latent features by using the patches and taking into account semantic relationships between patches of the grid-based scene representation, and d. at least one downstream DL module (6) for performing behavior planning and / or predicting the development of the traffic scenario based on the latent features; The method is characterized in that the perception module (1) is designed to distribute the scene-specific information to at least two semantic levels of the grid-based scene representation. The preprocessing component (2) is designed to maintain the distribution of the scene-specific information to at least two semantic levels when dividing into a plurality of blocks and form at least one type of sub-blocks, wherein a type of sub-blocks respectively comprises: a subset of the at least two semantic levels of the block defining the type, and The DL architecture (3) is designed to consider at least one type of semantic relationship between sub-patches when mapping the grid-based scene representation to the latent features.

10. The system (100) according to claim 9, It is characterized in that To map the grid-based scene representation to latent features, the DL architecture (3) includes the following components: a. at least one neural embedding network (4) for generating embeddings for one type of sub-block, and b. At least one processing network (5) for generating said latent features based on the embeddings thus generated.

11. The system (100) according to claim 10, It is characterized in that The DL architecture (3) includes a plurality of neural embedding networks (4) for generating embeddings for different types of sub-blocks, and the processing network (5) is designed to consider the embeddings of different types of sub-blocks when generating the latent features.

12. A system according to any one of claims 9 to 11, It is characterized in that The processing network is designed to take into account at least one further grid-based scene representation of at least one earlier time point, and in particular a time sequence of the grid-based scene representations, when generating the latent features for the grid-based scene representation.