Computer-implemented method and system for planning trajectory of host vehicle

Abstract

A computer-implemented method for planning a trajectory of a host vehicle in a traffic scene having further participants. According to the invention, an initial set Gi (0) of equivariant features and an initial set hi (0) of invariant features are generated for each participant by means of a DL-based initialization means on the basis of the trajectories of all participants. The features of the initial set Gi (0) and the initial set hi (0) are modified step by step in parallel but interacting branches by means of a neural network such that the isovariant features of the resulting set Gi (N) map the relationship between the participants of the traffic scene and target information specified for the host vehicle is taken into account when the features of the initial set G0 (0) are modified. Then, in a prediction step, at least one trajectory is predicted for all participants of the traffic scene by means of a DL-based trajectory decoder on the basis of the resulting sets Gi (N) of the isovariant features, respectively. Finally, a trajectory predicted for the host vehicle is selected as a planning result.

Description

Technical Field

[0001] The invention relates to a method for planning a trajectory of a vehicle in a traffic scene with at least one other participant. Computer-implemented methods and systems. Background Art

[0002] To plan safe and understandable maneuvers, automated vehicles need to anticipate how the current traffic scene will develop. This is particularly important when other participants in the scene, such as other vehicles, cyclists, and pedestrians, are also present. Therefore, attempts are made to predict the future behavior of all possible participants in the traffic scene so that this future behavior can be taken into account when planning the vehicle's maneuvers or trajectories. In recent years, the use of machine learning, particularly models based on deep learning (DL), has become the de facto standard for prediction.

[0003] In the context of the present invention, each trajectory is described based on a sequence of points in Euclidean space. Depending on the application, these are preferably points in 2D or 3D Euclidean space. In addition to the coordinates of the individual waypoints, the trajectory data may also include additional information, such as information describing the movement of the respective participant at the waypoints of the trajectory, in particular speed, orientation, etc.

[0004] Here, the planning for the vehicle is based on the following trajectory X i The following is done: the individual participants in the traffic scene have already traveled the trajectory at the planned time. In this case, not only the other participants in the traffic scene are taken into account, but also the trajectory X0 that the own vehicle has traveled.

[0005] In terms of validity and robustness, the predicted trajectory should be independent of the viewpoint of the traffic scene, that is, regardless of the perspective from which the traffic scene is observed. This requirement can be met by using a prediction model that is rotation-translation equivariant. In this type of prediction model, Euclidean transformations of the model input—translation, rotation, or reflection—affect the model output in the same way. In other words, when a Euclidean transformation is applied to the input variables, the predicted results are transformed in the same way, i.e., they are also rotated, translated, or mirrored.

[0006] EqMotion (https: / / arxiv.org / pdf / 2303.10876.pdf) involves a rotation-translation-equivariant prediction model that predicts the future course of motion at discrete time points. For example, EqMotion can also be used to predict the behavior of participants in traffic scenarios. However, EqMotion is limited to prediction; it is not designed for planning tasks. Summary of the Invention

[0007] The present invention proposes to use DL-based rotation-translation equivariance (roto-translations- ) model to predict the possible further development of a given traffic scenario. Trajectories are predicted for all participants in the traffic scenario in parallel. Consequently, when the model is trained accordingly, more complex interactions between the individual participants are also modeled, particularly interactions between the vehicle itself and other participants. The trajectory predicted for the vehicle in this way is then used as the basis for trajectory planning for the vehicle itself.

[0008] To this end, according to the invention, an initial set G of first features is generated for each participant in the traffic scene by means of a DL-based initialization component. i (0) and the initial set h of the second feature i (0) , more precisely based on the trajectory X that all participants have traveled at the planning time point i According to the present invention, the initialization component is designed and trained so that: relative to the trajectory X i The first characteristic of the Euclidean transformation is equivariance The second characteristic is invariant.

[0009] Therefore, the participant trajectory X i into a latent space, which can nevertheless still be assigned to the Euclidean initial space by means of each initial set G i (0) Represents a Cx2 or Cx3 tensor, where C is the feature embedding dimension and describes the number of 2D or 3D points in the latent space.

[0010] Translational equivariance is achieved by moving the coordinate origin to the center of all positions before applying the linear transformation. Rotational equivariance stems from the associativity of matrix multiplication, which also applies to processing with fully connected layers (FCL).

[0011] Then, according to the invention, the initial set G is modified step by step in parallel, but interacting branches by means of at least one neural network which is also translationally and rotationally equivariant. i (0) The characteristics and initial set h i (0) The characteristics of the obtained set G i (N) The equivariant feature maps the relationships between participants in the traffic scene.

[0012] That is, in these modifications, the initial set G i (0) Relative to trajectory X i Equivariance of Euclidean transformation and initial set h i (0) Relative to trajectory X i The invariance of the Euclidean transformation is also preserved.

[0013] According to the present invention, when modifying the initial set G0 of the vehicle (0) When determining the characteristics of the vehicle, target information predetermined for the vehicle is also taken into account.

[0014] Then, in the prediction step, the set G obtained based on the equivariant features is obtained by means of at least one DL-based trajectory decoder. i (N) Predict trajectories for all participants in a traffic scene More precisely, at least one trajectory is predicted for each participant.

[0015] Here, at least one trajectory is also predicted for the vehicle itself Since the set G0 on which the prediction is based (N) The relationships between all participants in the traffic scene are mapped, so that when making predictions for the vehicle itself, possible interactions with other participants in the traffic scene are also taken into account.

[0016] Finally, according to the present invention, from all the trajectories predicted for the vehicle Select at least one trajectory as the result of planning

[0017] That is, prediction and planning therefore proceed simultaneously / hand-in-hand, since the selection of a trajectory previously predicted for the host vehicle does not need to be learned, but can simply be made in a rule-based manner.

[0018] As mentioned above, behavior prediction and planning are based on the following trajectory X i The trajectories X of the participants in the traffic scene are usually determined based on the sensor data detected and processed by the perception module of the vehicle. i For example, track X i can be identified as a component of the environmental model.

[0019] In a preferred embodiment of the present invention, a graph neural network (GNN) based on an image is used to modify the initial set G i (0) The characteristics and initial set h i (0) The image-based neural network includes N fully connected layers FCL (fully connected layer). Therefore, two interacting processing branches are implemented, namely for the equivariant feature G i (e.g. vehicle position) and for the invariant feature h i (e.g., the distance between two vehicle positions).

[0020] GNNs are particularly well suited for identifying and predicting relationships and correlations between the past and future behaviors of various participants in a traffic scene.

[0021] It is important for the present invention that when modifying the initial set G0 of the vehicle (0) When considering the characteristics of the vehicle, the target information predefined for the vehicle is taken into account. The form in which the target information is predefined depends primarily on the application and the specific implementation. Thus, the target information can be predefined, for example, in the form of route information, which is determined by the navigation system based on the starting point, the destination point, and the map information. However, the target information can also consist solely of a direction (e.g., north) or a destination point (e.g., the next intersection in the direction of travel). Target information in the form of a predefined action, such as a left turn, is also conceivable.

[0022] In a particularly advantageous variant of the invention, a tensor L is generated based on the predetermined target information, which describes the target information in the form of a route and has the same dimension as the initial set G of equivariant features. i (0)That is, therefore, the number C of 2D or 3D waypoints of the target route is generated based on the target information, wherein the number C corresponds to the initial set G i (0) The feature embedding dimension C of the image is C. If the target information is given by more or fewer waypoints, the target information can be simply converted into a Cx2 or Cx3 tensor representation by interpolation.

[0023] It has been recognized that this type of tensor L can be used to feed target information into the neural network to modify the initial set G i (0) The characteristics and initial set h i (0) To this end, in each modification step, a tensor L is formed with the corresponding set G0 (l) The difference between (L-G0 (l) ). Then, gradually modify the initial set G0 (0) When the difference (L-G0 (l) ) is optimized. This optimization also affects the set G of other participants i (l) and the initial set h i (l) , since according to the invention the set G is modified step by step in parallel, but interacting branches i (l) and the initial set h i (l) .

[0024] In a preferred embodiment of the present invention, different possible developments of the traffic scenario are taken into account when predicting the behavior of the individual participants. This is called multimodal prediction. In this case, multiple trajectories are predicted for at least one participant in the traffic scene with the aid of one or more DL-based trajectory decoders. The plurality of trajectories represent different developments of the traffic scene.

[0025] In this context, it has proven to be particularly advantageous to have for each predicted trajectory At least one evaluation variable P is also predicted i k That is, then, when planning, one or more predicted trajectories for the vehicle are selected as planning results. When considering these evaluation parameters P i k .

[0026] In addition, multiple trajectories with different modalities The predicted evaluation parameter P i k It can represent the entry / occurrence probability of different modes. In the case of i k Can be used for quality estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In the following, exemplary embodiments and advantageous developments of the invention are explained in more detail with reference to the single FIGURE.

[0028] Figure 1 A method according to the present invention for planning a trajectory of a vehicle in a traffic scene with at least one other participant is shown. Schematic block diagram of a computer-implemented system 100.

[0029] In the following, according to Figure 1 The planning method according to the invention is also explained in more detail. DETAILED DESCRIPTION

[0030] In the only attached figure Figure 1 The system 100 according to the invention shown in FIG is a component of a vehicle, which is hereinafter referred to as the vehicle. The system 100 is used for trajectory planning of the vehicle. According to the invention, trajectory planning is based on the following trajectory X i : Each participant i of a given traffic scenario has walked through the trajectory at the time of the planning.

[0031] The path X is obtained by using the perception layer of the system 100. i 1, the perception layer is not shown in detail here. It is important to determine not only the trajectory X0 of the vehicle (i=0) in the traffic scene, but also the trajectory X0 of other participants (i≠0). i. For this purpose, the perception layer usually includes a plurality of different sensors, such as video sensors, radar sensors and / or lidar sensors, and possibly also inertial sensors, in order to aggregate scene-specific information. With the help of corresponding communication connections and access rights to the perception layer, the information detected by the vehicle's own sensor system can also be supplemented by information from information sources outside the vehicle (such as infrastructure sensors, GPS positioning, map information, etc.). Based on the scene-specific information aggregated in this way, the perception layer generates an environmental model of the traffic scene and its development up to the planned time point. In this case, based on the aggregated scene-specific information, object detection is performed in order to identify and locate the individual participants themselves and thus determine the trajectory X they have traveled. i 1, where each trajectory X i 1 is described here as a sequence of points in 2D Euclidean space.

[0032] Furthermore, system 100 includes a destination presetting module having an input interface for destination information related to the vehicle. In principle, this destination information can also be predefined in various forms, such as route information from a navigation system, direction information or a destination point to be driven to, or even a predefined behavior (e.g., left turn or straight ahead).

[0033] However, in the embodiment described herein, the target specification module serves more than just an input interface. Furthermore, it converts target information into a format suitable for further processing, as will be explained in greater detail below. The result of this conversion, i.e., the output of the target specification module, is indicated by 2 in the single figure.

[0034] The trajectory X obtained by the perception layer i 1 is supplied to a DL-based initialization component 3, which is designed such that it is based on the trajectories X of all participants. i 1 Generate an initial set G of the first feature for each participant i (0) and the initial set h of the second feature i (0) , where relative to trajectory X i The first characteristic of the Euclidean transformation is equivariant and the second characteristic is invariant.

[0035] Therefore, participant trajectory X i 1 is mapped into the latent space, but the latent space can still be assigned to the 2D Euclidean initial space in the way that each initial set G i (0)Represents a Cx2 tensor, where C is the feature embedding dimension, which corresponds to the number of 2D points in the latent space.

[0036] This initialization can be based on the trajectory X i 1 performs rotational and translational equivariant behavior prediction for each participant in the traffic scene.

[0037] The basic idea of ​​performing rotation and translation equivariance processing or operations on a sequence of points in Euclidean space is to move the observation point (Standpunkt) to the center, that is, to the center

[0038] In order to restore the original coordinate system after the transformation, the center of gravity Add again. Accordingly, each initial set G i (0) Can be described as:

[0039]

[0040] Here, the function φinit is implemented by initializing the fully connected layer FCL (fully connected layer) of the neural network of component 3 g , thereby ensuring rotation-translation equivariance. That is, FCL always performs a linear transformation of the corresponding input data and can be simply expressed in the form of matrix multiplication.

[0041] In addition, the invariant features of each participant i are initialized, that is, the following function: This function can be used to calculate the trajectory X from the obtained trajectory i The waypoints are derived from the corresponding speed ΔX of the participant i i and orientation angle (ΔX i τ ,ΔX i τ-1 )composition.

[0042]

[0043] As the function φinit g , by initializing the FCL of the neural network in component 3, the function φinit is also implemented h , thereby ensuring the invariance of the transformation.

[0044] In addition, the initialization component 3 can have learned to generate the initial set G i (0) and h i (0) Consider the relationship c between each participant i and jij , for example differences in distance, speed and / or orientation between the individual participants i and j.

[0045] As mentioned above, the target presetting module of the embodiment described here converts the predefined target information into a format suitable for further processing, that is, into the initial set G i (0) This enables the trajectory planning to be adapted to the target specification for the vehicle in a rotationally and translationally equivariant manner.

[0046] In the exemplary embodiment described here, the vehicle has a strategic planner or navigation system that, upon request, provides a rough route plan based on a road map, in this case in the form of an arbitrary number of 2D waypoints. This rough route plan is predefined as target information for the vehicle. A target predefined module maps the corresponding 2D waypoints to a set of C 2D waypoints by suitable interpolation, wherein the number C is represented by the tensor G. i (0) The set of 2D waypoints generated in this way forms a tensor L, which is also referred to below as high-level route L2.

[0047] Not only the advanced route L, but also the initial set G of equivariant features and invariant features i (0) and h i (0) Another neural network 4 is supplied to system 100. This neural network 4 is implemented as a GNN and includes N so-called feature update blocks, which are traversed sequentially. Each of these N feature update blocks is implemented via multiple operations. The learnable portion of each of these operations is implemented via FCLφ. However, the feature update blocks also reflect intentionally introduced decisions, such as the summation of all neighbors.

[0048] It should be noted here that in principle all operations of the feature update block can also be learned by GNN4.

[0049] Based on the configuration of GNN4, the initial feature G i (0) and h i (0) The gradual / continuous (sukzessiv) modification of the trajectory X as a whole i The Euclidean transformation of 1 is equivariant or invariant. The superscript (l) denotes the lth FCL or lth modification step of GNN4, respectively.

[0050] According to the present invention, GNN4 is designed and trained so that it modifies the initial set G in parallel but interacting branches. i (0) The characteristics and initial set h i (0) characteristics, more precisely modified so that the resulting set G i (N) The relationship between the participants of the traffic scene is mapped by the equivariant feature map and in addition at least in modifying the initial set G0 of the vehicle (0) The pre-given target information is considered when calculating the equivariant features.

[0051] In the embodiment described here, the equivariant feature G0 of the host vehicle is moved in the direction of the high-level route L in the latent space. (l) , more precisely by means of the individual FCLs of the GNN4, which are traversed in sequence and modify the results of the previous layers respectively. Therefore, when training the GNN4, in particular the vector (L-G0) between the high-level path L and the output of the corresponding FCL layer is optimized respectively. (l) ).

[0052] In each modification step, the equivariant feature G0 of the vehicle is firstly performed. (l) The movement towards the advanced route L is indicated by block 41. Only then are the interactions with the remaining participants considered in blocks 42 and 43. Thus, in block 42, for all participants, i.e. for all sets G i (l) The equivariant feature is modified. In block 43, for all participants, that is, for all sets h i (l) Modifications are made to the invariant features. These modifications are not performed independently of one another, but rather in parallel and in interaction with one another, as indicated by the corresponding arrows in block 4. This modification sequence results in the interaction between the participants taking precedence over adherence to the route plan of the host vehicle.

[0053] Before describing blocks 41, 42, and 43 in detail, it should be noted that all functions φ of these blocks are implemented by the FCL of GNN4, respectively. (l) , thereby ensuring the rotational-translational equivariance of the modification.

[0054] In block 41, only the equivariant feature G0 of the host vehicle is modified. (l) , since only the destination information or high-level route L is provided for the host vehicle. The modification in block 41 can be described as follows:

[0055]

[0056] Vector (L-G0 (l) ) forms the function φra (l) Or the input of the corresponding FCL, this vector points in the direction of the target information from the embedding of the equivariant features of the host vehicle in the form of a high-level route L.

[0057] The following equation (4) shows that the feature modification in block 41 satisfies the equivariance condition:

[0058]

[0059] In block 42, the equivariant features for all participants i - set G i (l) The modification of —— includes internal aggregation and adjacent aggregation. In internal aggregation, the invariant feature h based on the corresponding participant i is used. i (l) The calculated weight modifies the equivariant feature G of each participant i i (l) .

[0060]

[0061] Here, is the equivariant characteristic G of the corresponding participant i i (l) the center or middle position of the

[0062] In neighbor aggregation, we first define the edge weight e for each neighbor i, j ij , more precisely based on the relation feature c ij and equivariant features and invariant features, wherein K represents the number of different relations described by the relational features.

[0063]

[0064] Then, with the help of all its neighbors N i The weighted sum of modifies the i-th equivariant feature as follows:

[0065]

[0066] Finally, an equivariant nonlinear function is applied as described in EqMotion (https: / / arxiv.org / pdf / 2303.10876.pdf) to generate the input G for the next FCL. i (l+1) .

[0067] For the invariant features in block 43 - set h i (l)——, also refer to EqMotion (https: / / arxiv.org / pdf / 2303.10876.pdf). This modification can be described as follows:

[0068]

[0069] However, for the following predictions, only the equivariant features G of each participant i are used i (N) The resulting set.

[0070] In the embodiment described here, the system 100 comprises a DL-based trajectory decoder 5 in the form of K parallel-operating FCLs, which are respectively described as functions φ dec k Each of these FCLs φ dec k The prediction pattern k, ie the possible development of a given traffic scenario, is used. For this purpose, in a common prediction step, at least one trajectory is predicted for each participant i in the traffic scenario. More precisely, based on the equivariant characteristic G i (N) The resulting set is used for prediction.

[0071] Here, the superscript k represents the kth FCL or the kth predicted mode of the track decoder 5. The track can be described as follows Prediction:

[0072]

[0073] That is, the trajectory is realized for all participants in parallel and multimodally regression, that is, regressing multiple alternative future scenarios in parallel.

[0074] It is particularly advantageous if the trajectory decoder 5 is trained such that for each predicted trajectory At least one additional "waypoint" is predicted as an evaluation criterion. That is, each predicted trajectory Beyond the real prediction range T f The predicted trajectory is derived from additional “waypoints” Evaluation parameter P i k , by averaging the 2D coordinates in two dimensions:

[0075]

[0076] Parameter Pi k is an indicator of the entry probability for mode k.

[0077] Finally, the system 100 also comprises a selector member 7 designed such that it selects from the trajectory predicted for the host vehicle Select at least one trajectory as the planning result In this case, the selector element takes into account the predicted evaluation variable P0 k , the predicted evaluation variables can be interpreted as entry probabilities of different modes. The selector component described here selects the planning results in a rule-based manner. Make the following selections:

[0078]

[0079] The above-described embodiments illustrate that the present invention proposes a rotation-translation equivariant combination of trajectory prediction and trajectory planning. In the prediction step, trajectories are predicted for all participants in the traffic scene. This allows for modeling even more complex interactions between participants in the traffic scene. Actual trajectory planning then simply involves selecting at least one predicted trajectory for the vehicle itself. Rotation-translation equivariance means that the predicted or planned trajectory is transformed in the same way as the Euclidean transformation of the input trajectory.

[0080] Thereby, the amount of training data can be reduced, and the number of required learnable weights of the neural network can also be reduced.

Claims

1. A method for planning a trajectory of a vehicle in a traffic scene with at least one other participant (8), a computer-implemented method, - where the trajectory is described based on a sequence of points in Euclidean space, and -Wherein, based on the following trajectory X i (1) The plan is realized: each participant, i.e. the own vehicle and the at least one further participant, has traversed the trajectory at the time of the plan, It is characterized by: - Based on the trajectory X of all participants i (1) Generate an initial set G of the first features for each participant using the DL-based initialization component (3) i (0) and the initial set h of the second feature i (0) , where, relative to the trajectory X i (1), the first characteristic is equivariant and the second characteristic is invariant, - gradually modifying the initial set G in parallel but interacting branches by means of at least one neural network (4) i (0) The characteristics and the initial set h i (0) characteristics, more precisely modified so that these characteristics are relative to the trajectory X i The equivariance or invariance of the Euclidean transformation of (1) is preserved, and the resulting set G i (N) The equivariant feature map describes the relationship between the participants of the traffic scene, - At least modify the initial set G0 of the vehicle (0) The target information predetermined for the vehicle is also taken into account when considering the characteristics of - In the prediction step, by means of at least one DL-based trajectory decoder (5), the resulting set G is respectively based on the equivariant features i (N) Predict at least one trajectory for all participants in the traffic scene (6), and - Select at least one predicted trajectory for the host vehicle As a result of the planning (8).

2. The method according to claim 1, characterized in that The trajectory X of the participants in the traffic scene is obtained based on the sensor data detected and analyzed by the perception module of the vehicle. i (1).

3. The method according to any one of claims 1 or 2, characterized in that The initial set G is processed by means of at least one graph-based neural network GNN (4; 41, 42, 43) i (0) The characteristics and the initial set h i (0) The features are modified, and the image-based neural network GNN includes N fully connected layers FCL (fully connected layers).

4. The method according to any one of claims 1 to 3, characterized in that The destination information is specified in the form of route information, direction information, a destination point to be traveled to, and / or a specified behavior.

5. The method according to any one of claims 1 to 4, characterized in that Based on the predetermined target information, a tensor L(2) is generated, which describes the target information in the form of a route and has the same dimension as the initial set G of equivariant features. i (0) The dimensions are the same.

6. The method according to claim 5, characterized in that When modifying the initial set G0 of the vehicle (0) The predetermined target information is taken into account when considering the characteristics of the tensor L and the corresponding set G0 in each modification step. (l) The difference between (L-G0 (l) ), and gradually modify the initial set G0 (0) When optimizing the difference (L-G0 (l) )(Block 41).

7. The method according to any one of claims 1 to 6, characterized in that Predicting multiple trajectories of different modalities for at least one participant in the traffic scene with the aid of the at least one DL-based trajectory decoder (5) The different modes represent different developments of the traffic scenario (multimodal prediction).

8. The method according to any one of claims 1 to 7, characterized in that By means of the at least one DL-based trajectory decoder (5), for each predicted trajectory Predict at least one evaluation parameter And considering the predicted evaluation parameters In the case of, selecting at least one trajectory predicted for the vehicle As a result of the planning (8).

9. The method according to claims 7 and 8, characterized in that Multiple trajectories with different modalities Evaluation parameters predicted together Represents the entry probability of the different modes and the predicted trajectory for the vehicle is used for quality estimation in the case of 10. A method for planning a trajectory of a vehicle in a traffic scene with at least one other participant (8), a computer-implemented system (100), in particular for executing the method according to any one of claims 1 to 9, comprising at least: - a target presetting module, configured to preset target information for the vehicle; A perception layer for determining the trajectory X of each participant, ie the own vehicle and the at least one further participant, based on the aggregated, scene-specific information i (1); - a DL-based initialization component (3) designed such that it is based on the trajectories X of all participants i (1) Generate the initial set G of the first feature for each participant i (0) and the initial set h of the second feature i (0) , where, relative to the trajectory X i a Euclidean transformation of , wherein the first characteristic is equivariant and the second characteristic is invariant; - a neural network (4; 41, 42, 43) designed to: ■ Make the neural network modify the initial set G in parallel, but interacting branches i (0) The characteristics and the initial set h i (0) features such that these features are relative to the trajectory X i The equivariance or invariance of the Euclidean transformation of (1) is preserved, and the resulting set G i (N) The equivariant feature map represents the relationship between the participants of the traffic scene; and ■ Make the neural network at least modify the initial set G0 of the vehicle (0) The predetermined target information is also taken into account when considering the characteristics of - at least one KI-based trajectory decoder (5), which is designed so that in the prediction step, the at least one KI-based trajectory decoder is based on the obtained set G of equivariant features i (N) Predict at least one trajectory for each participant in the traffic scene (6); and - a selector component (7) which selects the trajectory predicted for the host vehicle As a result of the planning (8).

11. The system (100) according to claim 10, characterized in that The neural network is implemented in the form of a graph neural network GNN (4; 41, 42, 43), which includes N fully connected layers FCL (fully connected layers).

12. The system (100) according to any one of claims 10 or 11, characterized in that The target pre-given module is designed so that the target pre-given module generates a tensor L(2) based on the pre-given target information, the tensor describes the target information in the form of a route, and the dimension of the tensor is the same as the set G0 generated for the vehicle of the equivariant feature. (l) have the same dimensions as The neural network (4; 41) is trained so that the initial set G0 is modified step by step. (0) When optimizing the tensor L and the corresponding set G0 (l) The difference between (L-G0 (l) ).

13. The system (100) according to any one of claims 10 to 12, characterized in that K DL-based trajectory decoders operating in parallel are provided, each of which predicts at least one trajectory for all participants in the traffic scene, more precisely for each of K possible developments of the traffic scene. (6) (Multimodal prediction).

14. The system (100) according to any one of claims 10 to 13, characterized in that The at least one DL-based trajectory decoder (5) is trained so that the DL-based trajectory decoder is trained for each predicted trajectory. (6) Predict at least one evaluation parameter And the selector component (7) is designed so that when selecting the planning result (8) Consider the predicted evaluation parameters 15. A vehicle having a computer-implemented system (100) according to any one of claims 10 to 14 for trajectory planning, in particular for trajectory planning by means of a method according to any one of claims 1 to 9.

Images