Method for trajectory prediction and method for automatically driving a vehicle

Abstract

The present invention relates to a method for trajectory prediction of vehicles (F1-Fn) in the environment of the ego-vehicle. According to the invention, the interaction degrees (I1-Im) between the vehicles (F1-Fn) are determined using a machine-trained attention-based interaction algorithm (3). Using the interaction algorithm (3), individual vehicles (F1-Fn) are identified as relevant for trajectory prediction from the set of vehicles (F1-Fn) present in the environment of the ego-vehicle and are selected for trajectory prediction if their respective interaction degrees (I1-Im) with at least one of the vehicles (F1-Fn) for which a trajectory (T) is predicted exceed a predefined limit value. In a subsequent learning step, a trajectory prediction algorithm (5) is trained with the vehicles (F1-Fn) selected as relevant for the trajectory prediction, and the trajectory prediction performed using the trajectory prediction algorithm (5) is implemented for the vehicles (F1-Fn) selected as relevant for the trajectory prediction. Furthermore, the present invention relates to a method for automatically driving an ego-vehicle.

Description

Technical Field

[0001] The present invention relates to a method for predicting the trajectory of a vehicle in the surrounding environment of a host vehicle as described in the preamble of claim 1.

[0002] Furthermore, the present invention relates to a method for automatically driving a host vehicle.

Background Art

[0003] From the prior art, generally, automated vehicles, for example highly automated vehicles or autonomous vehicles, are known. In order to be able to meet the safety requirements imposed on such vehicles, it is necessary to predict the future movements of the vehicles involved in traffic. This is also referred to as trajectory prediction. In this trajectory prediction, an approach based on machine learning methods is known.

[0004] Hereinafter, the term "scene" means a surrounding environment situation in which at least one additional vehicle may exist, detected using a surrounding environment detection unit of a vehicle. In order to describe such a scene, map data of a digital road map can also be used additionally.

[0005] For example, different from a rasterization approach such as a CNN-based approach (CNN: Convolutional Neural Network), an approach for trajectory prediction that does not perform rasterization of a scene is based on a so-called graph neural network (GNN: Graph Neural Network), or self-attention, also referred to hereinafter as an attention-based approach. In order to model the social interaction between vehicles, a graph is formed in these approaches. The nodes of the graph are the encodings or codings of the vehicles in the scene. The nodes are connected to each other via edges.

[0006] Approaches exist that connect all vehicles in a scene with edges so that a complete graph, also called a fully-connected graph, is formed, which means that all vehicles are selected as being included or relevant in the trajectory prediction of the target vehicle.

[0007] In another approach, the edges of the graph are formed according to the distance between individual vehicles, thus providing a distance-based pre-selection of the vehicles to be considered for the prediction of the target vehicle.

[0008] Y. Ma et al., "Traffic Prediction: Trajectory Prediction for Heterogeneous Traffic-Agents; In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, 2019, pp. 6120-6127, doi: 10.1609 / aaai.v33i01.33016120," describes a method for predicting future trajectories of road users for autonomous vehicle navigation. A real-time traffic prediction algorithm based on so-called long short-term memory (LSTM) is used. This includes an instance level for learning the behavior and interactions of instances. Furthermore, the real-time traffic prediction algorithm also includes a category level for learning the similarity of instances of the same type to refine predictions. Summary of the Invention [Problem to be solved by the invention]

[0009] The object of the present invention is to provide a novel method for predicting the trajectory of a vehicle in its surrounding environment and a novel method for automatically driving the vehicle. [Means for solving the problem]

[0010] According to the present invention, the above problem is solved by a method for trajectory prediction having the constituent elements described in claim 1 and a method for automatically driving a host vehicle having the constituent elements described in claim 9.

[0011] Advantageous embodiments of the present invention are the subject matter of the dependent claims.

[0012] In a method for predicting the trajectory of a vehicle in the surrounding environment of a host vehicle, according to the present invention, a machine-trained attention-based interaction algorithm is used to determine the degree of interaction between vehicles. That is, the degree of interaction is determined for both the interaction between vehicles in the surrounding environment of the host vehicle and the interaction between the host vehicle and the vehicles in the surrounding environment of the host vehicle. Further, using the interaction algorithm, individual vehicles are identified as relevant to trajectory prediction from the set of vehicles existing in the surrounding environment of the host vehicle, and when the degree of interaction of each of the individual vehicles with at least one vehicle whose trajectory is predicted exceeds a predetermined limit value, it is selected for that trajectory prediction. In a subsequent learning step, the trajectory prediction algorithm is trained using the vehicles selected as vehicles relevant to trajectory prediction, particularly trained using only the vehicles selected as vehicles relevant to trajectory prediction, and the trajectory prediction executed using the trajectory prediction algorithm is performed only for the vehicles selected as vehicles relevant to trajectory prediction, particularly only for the vehicles selected as vehicles relevant to trajectory prediction.

[0013] In particular, this method constitutes a so-called two-step method for predicting the trajectory of a vehicle. In this two-step method, first, interaction-based vehicle pre-selection is performed, and subsequently, trajectory prediction is performed using a reduced number of vehicles.

[0014] Learning-based trajectory prediction algorithms, for example, include map data of high-resolution digital road maps and are often based on a graph structure. In this graph structure, typically, other vehicles detected in the surrounding environment of the host vehicle form nodes in the graph. Therefore, depending on the number of vehicles in the scene, the complexity of such trajectory prediction algorithms is extended. This results in a strong runtime dependency, especially when map information also has to be queried for all vehicles in the scene. Furthermore, a large number of irrelevant vehicles in the scene make the learning process of the trajectory prediction algorithm difficult, for example, resulting in insufficient prediction results and insufficient convergence behavior. For example, vehicles at a far distance do not have a measurable impact on the trajectory prediction of the vehicle to be predicted. Therefore, when all agents or vehicles in the scene are selected, the trajectory prediction algorithm needs to independently identify which vehicles have a beneficial added value for the prediction and which do not.

[0015] In contrast, in the present method, an initial pre-selection of vehicles relevant to the trajectory prediction algorithm is realized using an interaction algorithm. In a particularly advantageous method for determining the interaction, since map data of a high-resolution digital road map is not required, the interaction algorithm is characterized by low complexity. Therefore, especially in the case of complex and sometimes map-based trajectory prediction algorithms, the complexity can be reduced by selectively choosing vehicles. In a graph-based approach, the number of nodes in the graph can be reduced, thus also shortening the execution time.

[0016] Unlike distance-based selection, the pre-selection of relevant vehicles is not performed according to the distance of the relevant vehicles, but according to the degree of interaction between the relevant vehicles and the vehicle to be predicted. Therefore, for example, it can be considered that the future trajectory of a certain vehicle is strongly affected by a vehicle traveling ahead at a certain distance. In contrast, there is little influence from a vehicle behind traveling in the same direction.

[0017] Therefore, by this method, compared with the approach without prior selection, an improvement in the execution time of the final trajectory prediction algorithm can be achieved. Also, by the prior selection of relevant vehicles, the training time of the final trajectory prediction algorithm can be shortened. Furthermore, compared with the approach without prior selection and compared with the approach with distance-based prior selection, an improvement in the quality of the final trajectory prediction can be realized. The metric for the interaction between vehicles can also be additionally used for other tasks, such as affecting other road users.

[0018] Due to the prior selection and the accompanying reduction in the number of selected vehicles, even a very complex trajectory prediction algorithm can be used with low processing costs and short execution times. The trajectory prediction algorithm used here is, for example, - C. Tang and R. R. Salakhutdinov: “Multiple Futures Prediction”; In: Advances in Neural Information Processing Systems, 2019, vol. 32, - T. Salzmann, B. Ivanovic, P. Chakravarty and M. Pavone: “Trajectron++: Dynamically-Feasible Trajectory Forecasting with Heterogeneous Data”; In: Computer Vision-ECCV 2020, Cham, 2020, pp. 683 - 700, - M. Liang et al.: “Learning Lane Graph Representations for Motion Forecasting”; In: Computer Vision-ECCV 2020, Cham, 2020, pp. 541 - 556, - S. Khandelwal, W. Qi, J. Singh, A. Hartnett and D. Ramanan: “What-If Motion Prediction for Autonomous Driving” 2020, and / or - Y.Ma, X.Zhu, S.Zhang, R.Yang, W.Wang and D.Manocha: “TrafficPredict:Trajectory Prediction for Heterogeneous Traffic-Agents”; In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019,2019,pp.6120-6127,doi:10.1609 / aaai.v33i01.33016120 It is formed according to

[0019] In one possible configuration of the method, the degree of interaction between the vehicles is determined pairwise for each two vehicles.

[0020] In one possible configuration of the method, during machine training of the interaction algorithm, the dependencies between vehicles are learned pairwise, and the interaction algorithm is used to weight the extent to which the vehicles influence each other pairwise to form an interaction degree, so that during subsequent operation of the interaction algorithm, the interactions between the vehicles can be easily and reliably determined and the interaction degree can be determined.

[0021] In one possible configuration of the method, machine training of the interaction algorithm generates trajectory predictions that are used to train the interaction algorithm, thereby achieving effective training of the interaction algorithm and achieving very high accuracy and reliability of the interaction algorithm in determining the interaction degree.

[0022] In one other possible configuration of the method, the generated trajectory predictions are used in machine training of the interaction algorithm for implicit learning of the degree of interaction.

[0023] In one possible implementation of the method, during operation of the machine-trained interaction algorithm, all vehicles in the scene are encoded using long short-term memory (also called long short-term memory), forming the nodes of a complete graph, also called a fully connected graph. Because the graph's principle behavior is encoded with corresponding weights, the use of long short-term memory achieves the memory of previous experiences, and thus long-lasting short-term memory. Here, the complete graph can be used to fully represent the relationships between vehicles.

[0024] In one possible configuration of the method, the nodes are connected to each other via edges and the distance between the vehicles is used as an edge feature, so that, for example, the distance between the vehicles can also be taken into account when determining each interaction degree.

[0025] In another possible configuration of the method, the trajectory prediction algorithm uses map data from a high-resolution digital road map for trajectory prediction. By using such a high-resolution digital road map, the current location and past movements of vehicles detected and located in the scene can be stored. Here, the road map can include road topology, traffic signs, traffic signals (e.g., traffic light signals), crosswalks, and other information, thus achieving highly accurate and reliable trajectory prediction.

[0026] In the method according to the invention for automatically driving an ego-vehicle, in the method for predicting a trajectory of a vehicle in its surroundings according to the above, the trajectory of the vehicle in its surroundings is predicted, and the predicted trajectory is taken into account in the automated lateral and / or longitudinal control of the ego-vehicle during the automated driving of the ego-vehicle. By using the method for predicting a trajectory of a vehicle, particularly reliable and safe automated driving, such as highly automated or autonomous driving, of the ego-vehicle can be achieved.

[0027] Hereinafter, an embodiment of the present invention will be described in detail with reference to the drawings. [Brief explanation of the drawings]

[0028]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

DETAILED DESCRIPTION OF THE INVENTION

[0029] In any of the figures, corresponding parts are denoted by the same reference numerals.

[0030] FIG. 1 shows a block diagram of a possible embodiment of an apparatus 1 for predicting the orbits of a plurality of vehicles F1 to Fn in the surrounding environment of the host vehicle.

[0031] The data memory 2 stores data describing the scene in the surrounding environment of the host vehicle. These data include surrounding environment data detected by a surrounding environment sensor system (not shown in detail for the host vehicle, other vehicles and / or traffic infrastructure), and map data KD of a high-resolution digital road map. Here, in particular, the surrounding environment data also includes data regarding the vehicles F1 to Fn.

[0032] In order to minimize the processing effort when predicting the trajectories of vehicles F1 to Fn in the surrounding environment of the host vehicle, initial interactions are determined between the vehicle F1 to Fn whose trajectory is to be determined and all other vehicles F1 to Fn in the scene, i.e., including the host vehicle.

[0033] This determination is made by a machine-trained attention-based interaction algorithm 3, which determines multiple interaction degrees I1 to Im between the vehicles F1 to Fn, in particular between the vehicle F1 to Fn whose trajectory T is predicted and the vehicles F1 to Fn present in its surrounding environment. The interaction degrees I1 to Im between the vehicles F1 to Fn are determined for each pair of two vehicles F1 to Fn (for every two vehicles F1 to Fn).

[0034] This interaction algorithm 3 is characterized by its short execution time and low complexity, and focuses only on the interactions between vehicles F1 to Fn. High-resolution digital road map data KD is not required to determine the interaction degrees I1 to Im. Here, the interaction degrees I1 to Im represent so-called attention weights, which are formed by the so-called self-attention 3.1 of the attention-based interaction algorithm 3.

[0035] In particular, when determining the interaction degrees I1 to Im, a trajectory prediction TP′ is also generated that is used to train the interaction algorithm 3. Here, the trajectory prediction TP′ is used in machine training of the interaction algorithm 3, for example, in implicit learning of the interaction degrees I1 to Im.

[0036] Note that the interaction degrees I1-Im are not "labelable" and therefore cannot be explicitly learned from the data. Therefore, an attention-based interaction algorithm 3 is used that implicitly learns the interaction degrees I1-Im between vehicles F1-Fn. In contrast, the training objective, i.e., the predicted trajectory T, is "labelable." However, the interactions between vehicles F1-Fn are implicitly learned.

[0037] Specifically here, the interaction algorithm 3 learns to utilize the dependencies for each pair among the individual vehicles F1 to Fn, in particular, to determine the interaction degrees I1 to Im. That is, the learning of the dependencies for each pair among the vehicles F1 to Fn is carried out. Here, the attention mechanism utilized "weights" how strongly each of the vehicles F1 to Fn affects each other, for each pair respectively. The attention weights utilized here take values in the range of, for example, 0.0 to 1.0. Therefore, in order to predict the trajectories of the selected vehicles F1 to Fn and / or the host vehicle, it is possible to learn from the population of the dataset how important each of the surrounding vehicles F1 to Fn in the sequence or scene is.

[0038] After the training of the interaction algorithm 3 is completed, this interaction algorithm 3 is used particularly in the so-called Forward-Pass, in which case the trajectory predicted by the interaction algorithm 3 plays a subordinate role. Rather, here, the attention weights of the vehicles F1 to Fn to be predicted, which form the interaction degrees I1 to Im, are extracted. These indicate, for each of the other vehicles F1 to Fn in the sequence, how important each of those other vehicles F1 to Fn is for the prediction of the selected vehicles F1 to Fn and / or the host vehicle.

[0039] Based on the respectively determined interaction degrees I1 to Im, the selection unit 4 performs vehicle pre-selection A(F1 to Fn). That is, the individual vehicles F1 to Fn from the set of vehicles F1 to Fn existing in the surrounding environment of the host vehicle are identified as vehicles related to the trajectory prediction, and when the respective interaction degrees I1 to Im between that individual vehicle F1 to Fn and the other vehicles F1 to Fn exceed a predetermined threshold value, they are selected for that trajectory prediction. Here, for each of the vehicles F1 to Fn for which the trajectory T is predicted, an individual selection of the related vehicles F1 to Fn is performed. This is realized because the interaction degrees I1 to Im are formed in pairs. For example, only a certain number of the vehicles F1 to Fn having the highest attention weights are selected, or all of the vehicles F1 to Fn having an attention weight of 0.3 or more are selected.

[0040] The arbitrary complex trajectory prediction algorithm 5 is used only when vehicle pre-selection A (F1 to Fn) is performed. This trajectory prediction algorithm 5 also uses map data KD from a high-resolution digital road map, and finally predicts the trajectory T of at least one vehicle among vehicles F1 to Fn in the surrounding environment of the host vehicle based on the pre-selected vehicles F1 to Fn and the map data KD.

[0041] Both the training of the trajectory prediction algorithm 5 and the operation of the trajectory prediction algorithm 5 in an automatically drivable host vehicle using the finally predicted trajectory T are performed only based on vehicle pre-selection A (F1 to Fn) with the reduced number of interacting vehicles F1 to Fn. With vehicle pre-selection A (F1 to Fn), the final trajectory prediction algorithm 5 can focus only on relevant or interacting vehicles F1 to Fn, and in particular, a data-driven learning process is facilitated by better convergence. When the final trajectory prediction algorithm 5 additionally uses map data KD from a digital road map, this process is also simplified. The reason is that so-called "queries" from the road map, such as the distance to the next stop point or the allowable maximum speed, can become very complicated in terms of time, so by pre-selecting relevant vehicles F1 to Fn, the number of "queries" can be significantly reduced.

[0042] FIG. 2 shows a block diagram of a possible embodiment of an apparatus for forming the interaction algorithm 3 according to FIG. 1. Here, the interaction algorithm 3 is designed as an attention-based trajectory prediction algorithm to be implemented. In a manner not shown in detail, other configurations of the interaction algorithm 3 are possible as long as the attention-based interaction degrees I1 to Im are determined.

[0043] In the input encoder 6, all vehicles F1 to Fn are encoded using so-called LSTM encoders 6.1 to 6.n equipped with long- and short-term memory, and during this encoding, the nodes h1 of the complete graph G 0 ~h n 0That is, the resulting encoding is a so-called fully connected graph G with a node h1 0 ~h n 0 Node h1 0 ~h n 0 is the edge 1,2 , e 1,n , e 2,1 , e 2,n , e n,1 , e n,2 The distance between the vehicles F1 to Fn is calculated as an additional edge feature, i.e., the edge e 1,2 , e 1,n , e 2,1 , e 2,n , e n,1 , e n,2 is used to form

[0044] Here, the graph G is divided into multiple layers L1 to L2 in the interaction module 7 that forms the core of the interaction algorithm 3. g Furthermore, the interaction module 7 The so-called parallel self-attention heads SAL1 to SAL2 are required for this method. h Using these, so-called multi-head self-attention is performed following Ashish Vaswani et al.: “Attention Is All You Need” (https: / / arxiv.org / abs / 1706.03762).

[0045] Self-Attention Head SAL1~SAL h The layers L1 to L g The vector v obtained using (Lg) are supplied, and in this case, based on the combination unit 7.1, the self-attention heads SAL1 to SAL h The output results are combined and linearized using the function , and the resulting features a1 to a are formed as latent encodings. x are determined, and their features a1 to a xis updated based on the interaction degrees I1~Im and is also used for training. Subsequently, the features a1~a x are decoded into the trajectory T in order to realize the implicit learning of the interaction degrees I1~Im. That is, the dependencies and / or weights for each pair among the vehicles F1~Fn within the scene, and thus the interaction degrees I1~Im among the agents, i.e., among the vehicles F1~Fn, are learned in the interaction module 7. Each of the interaction degrees I1~Im is formed as the weight for each pair as a result of, for example, the so-called softmax function according to Ashish Vaswani et al.: "Attention Is All You Need" (https: / / arxiv.org / abs / 1706.03762). Subsequently, the interaction degrees I1~Im are further used to form the features a1~a x In particular, the interaction degrees I1~Im are stored in an nxn matrix M.

[0046] The output decoder 8 generates the trajectory prediction TP′ for the vehicles F1~Fn from the interaction degrees I1~Im using so-called linear residual decoders 8.1~8.z, but this trajectory prediction TP′ is only used for the training of the interaction algorithm 3. When applied to the trajectory prediction algorithm 5, this trajectory prediction TP′ is not used, and the vehicle pre-selection A(F1~Fn) is performed using only the attention weights, i.e., the interaction degrees I1~Im. The residual layer is understood to be a layer consisting of multiple neuron layers. Different from the conventional multi-layer perceptron (abbreviation: MLP), a so-called "skip connection" is introduced into the residual layer. Thereby, it is realized to add the encoding of the first layer to the encoding of the last layer by linear operation. Thereby, during training, the linear backpropagation of the error is realized. This shows good results especially in deep neural networks.

[0047] Figures 3 to 5 show three established metrics used to determine the quality of the trajectory prediction.

[0048] As shown in FIG. 3, the metrics include a change ΔminADE in the minimum average displacement error, which is included according to a fixed number C of vehicles F1 to Fn for each scene. This scene refers to the scene related to the trajectory prediction obtained based on the vehicle pre-selection A (F1 to Fn) by the trajectory prediction algorithm 5 described in FIGS. 1 and 2 (hatched bar), and the scene related to the trajectory prediction algorithm in which the number C of vehicles F1 to Fn is restricted using a distance-based method (non-hatched bar). The minimum average displacement error represents how far, on average, each calculated position of each trajectory T is from the true position.

[0049] Furthermore, as shown in FIG. 4, the metrics include a change ΔminFDE in the minimum final displacement error, which is included according to a fixed number C of vehicles F1 to Fn for each scene. This scene refers to the scene related to the trajectory prediction obtained based on the vehicle pre-selection A (F1 to Fn) by the trajectory prediction algorithm 5 described in FIGS. 1 and 2 (hatched bar), and the scene related to the trajectory prediction algorithm in which the number C of vehicles F1 to Fn is restricted using a distance-based method (non-hatched bar). Here, the minimum final displacement error represents the deviation of the prediction from the true trajectory T at each last prediction step.

[0050] Furthermore, as shown in FIG. 5, the metrics include a change ΔMR in the error classification rate, which is included according to a fixed number C of vehicles F1 to Fn for each scene. This scene refers to the scene related to the trajectory prediction obtained based on the vehicle pre-selection A (F1 to Fn) by the trajectory prediction algorithm 5 described in FIGS. 1 and 2 (hatched bar), and the scene related to the trajectory prediction algorithm in which the number C of vehicles F1 to Fn is restricted using a distance-based method (non-hatched bar).

[0051] For all metrics, lower values are better.

[0052] Based on these metrics, a comparison is made of how the interaction-based vehicle preselection A (F1-Fn) described above performs compared to distance-based (Euclidean) selection for the established final trajectory prediction algorithm 5 described above.

[0053] Each 0 value is given by the result of the final trajectory prediction algorithm 5, trained and tested on the entire dataset. Only deviations from this baseline are represented. Here, the number C represents a fixed number for which vehicles F1 to Fn are reduced in each scene.

[0054] These metrics clearly show that for each value of C, using the trajectory prediction algorithm 5 based on the above-mentioned interaction-based preselection A (F1 to Fn) leads to better results than using distance-based selection. That is, an improvement in the quality of the final trajectory prediction can be achieved compared to the approach using distance-based preselection.

[0055] The minimum mean displacement error change ΔminADE and the minimum final displacement error change ΔminFDE show improved performance compared to the final trajectory prediction algorithm 5 trained on the entire dataset due to the reduced complexity of the number C of 5, 7, and 9 vehicles F1-Fn, as can be seen from the negative values. This means that an improved quality of the final trajectory prediction can be achieved compared to the approach without pre-selection.

[0056] Furthermore, the interaction-based pre-selection A(F1-Fn) can achieve a faster execution time for the final trajectory prediction algorithm 5 compared to the trajectory prediction algorithm 5 based on the entire scene including all vehicles F1-Fn. [Prior art documents] [Non-patent literature]

[0057] [Non-Patent Document 1] Ye.

Claims

1. A method for predicting the trajectories of vehicles (F1 to Fn) in the surrounding environment of a host vehicle, comprising: - Using a machine-trained attention-based interaction algorithm (3), an interaction degree (I1 to Im) between the vehicles (F1 to Fn) is determined; - Using the interaction algorithm (3), each individual vehicle (F1 to Fn) is identified as being relevant to the trajectory prediction from the set of vehicles (F1 to Fn) present in the surrounding environment of the host vehicle, and when the respective interaction degree (I1 to Im) with at least one of the vehicles (F1 to Fn) for which a trajectory (T) is predicted exceeds a predetermined threshold value, it is selected for the trajectory prediction; - In a subsequent learning step, a trajectory prediction algorithm (5) is trained using the vehicles (F1 to Fn) selected as being relevant to the trajectory prediction; - The trajectory prediction executed using the trajectory prediction algorithm (5) is performed on the vehicles (F1 to Fn) selected as being relevant to the trajectory prediction; wherein the interaction degree (I1 to Im) between the vehicles (F1 to Fn) is determined pairwise for each pair of two of the vehicles (F1 to Fn). A method, characterized by the above.

2. A method for predicting the trajectories of vehicles (F1 to Fn) in the surrounding environment of a host vehicle, comprising: - Using a machine-trained attention-based interaction algorithm (3), an interaction degree (I1 to Im) between the vehicles (F1 to Fn) is determined; - Using the interaction algorithm (3), each individual vehicle (F1 to Fn) is identified as being relevant to the trajectory prediction from the set of vehicles (F1 to Fn) present in the surrounding environment of the host vehicle, and when the respective interaction degree (I1 to Im) with at least one of the vehicles (F1 to Fn) for which a trajectory (T) is predicted exceeds a predetermined threshold value, it is selected for the trajectory prediction; - In a subsequent learning step, a trajectory prediction algorithm (5) is trained using the vehicles (F1 to Fn) selected as being relevant to the trajectory prediction; - The trajectory prediction executed using the trajectory prediction algorithm (5) is performed on the vehicles (F1 to Fn) selected as being relevant to the trajectory prediction; In the machine training of the interaction algorithm (3), the dependencies between the vehicles (F1 to Fn) are learned pairwise, and the degree of mutual influence of the vehicles (F1 to Fn) is weighted pairwise to form the interaction degrees (I1 to Im) using the interaction algorithm (3). A method characterized by this. **Claim 3**: A method for predicting the trajectories of vehicles (F1 to Fn) in the surrounding environment of a host vehicle, comprising: - Using a machine-trained attention-based interaction algorithm (3) to determine the interaction degrees (I1 to Im) between the vehicles (F1 to Fn); - Using the interaction algorithm (3), each vehicle (F1 to Fn) is identified as relevant to the trajectory prediction from the set of vehicles (F1 to Fn) present in the surrounding environment of the host vehicle, and when the respective interaction degrees (I1 to Im) with at least one of the vehicles (F1 to Fn) for which a trajectory (T) is predicted exceed a predetermined threshold value, it is selected for the trajectory prediction; - In a subsequent learning step, a trajectory prediction algorithm (5) is trained using the vehicles (F1 to Fn) selected as relevant to the trajectory prediction; - The trajectory prediction executed using the trajectory prediction algorithm (5) is performed on the vehicles (F1 to Fn) selected as relevant to the trajectory prediction; In the machine training of the interaction algorithm (3), trajectory predictions are generated for use in training the interaction algorithm (3). A method characterized by this. **Claim 4** The method according to claim 3, characterized in that the generated trajectory predictions are used in the machine training of the interaction algorithm (3) for implicit learning of the interaction degrees (I1 to Im). **Claim 5**: A method for predicting the trajectories of vehicles (F1 to Fn) in the surrounding environment of a host vehicle, comprising: - Using a machine-trained attention-based interaction algorithm (3) to determine the interaction degrees (I1 to Im) between the vehicles (F1 to Fn); - Using the interaction algorithm (3), each of the vehicles (F1 to Fn) is identified as relevant to the trajectory prediction from the set of the vehicles (F1 to Fn) present in the surrounding environment of the host vehicle, and when each of the interaction degrees (I1 to Im) with at least one of the vehicles (F1 to Fn) for which a trajectory (T) is predicted exceeds a predetermined limit value, it is selected for the trajectory prediction. - In a subsequent learning step, a trajectory prediction algorithm (5) is trained using the vehicles (F1 to Fn) selected as relevant to the trajectory prediction. - The trajectory prediction executed using the trajectory prediction algorithm (5) is performed on the vehicles (F1 to Fn) selected as relevant to the trajectory prediction. During the operation of the interaction algorithm (3) trained by the machine, all the vehicles (F1 to Fn) in the scene are encoded using long- and short-term memory, and during the encoding, nodes (h 1 0 ~h n 0 ) of a complete graph (G) are formed A method, characterized by the above.

6. The node (h 1 0 ~h n 0 ) is interconnected via edges (e 1,2 , e 1,n , e 2,1 , e 2,n , e n,1 , e n,2 ), and the distance between the vehicles (F1 to Fn) is used as an edge feature, the method according to claim 5, characterized in that.

7. The method according to claim 1, characterized in that map data (KD) of a high-resolution digital road map is used for the trajectory prediction by the trajectory prediction algorithm (5).

8. A method for automatically driving a host vehicle, comprising: - In the method according to any one of claims 1 to 7, a trajectory (T) of vehicles (F1 to Fn) in the surrounding environment of the host vehicle is predicted. - The predicted trajectory (T) is taken into account in the automated lateral and / or longitudinal control of the host vehicle during the automatic driving of the host vehicle. A method, characterized by the above.

Images