Track planning method, related device, vehicle and storage medium

Abstract

The invention discloses a trajectory planning method, a related device, a vehicle and a storage medium, and the method comprises the steps: obtaining vehicle agent information and road environment information, the vehicle agent information carrying vehicle driving information of a target vehicle and vehicle driving information of other vehicles, the other vehicles are vehicles in the same driving area as the target vehicle; based on the vehicle agent information and the road environment information, determining a plurality of first driving tracks of the target vehicle in a future preset time period and a plurality of second driving tracks of other vehicles in the future preset time period; and a trained trajectory interaction module is adopted to process the plurality of first driving trajectories and the plurality of second driving trajectories to obtain a target prediction trajectory of the target vehicle, and the trajectory interaction module is constructed based on a game methodology. According to the scheme, the accuracy of target vehicle trajectory planning can be improved.

Description

Technical Field

[0001] The present application relates to the field of vehicle technology, and in particular to a trajectory planning method and related devices, vehicles, and storage media. Background Art

[0002] In autonomous vehicles (AVs), it is a major challenge to reasonably interact with other vehicles in complex traffic environments. This task requires considering the interactive impact of the ego-vehicle's decision-making behavior on other vehicles and accurately predicting their future behaviors to achieve effective decision-making.

[0003] The existing mainstream method uses a data-driven neural network model to predict the behavior of other vehicles to obtain their future trajectories, and then plans the trajectory of the vehicle based on the future trajectories of other vehicles. However, this solution does not take into account that the vehicle itself will often affect the trajectories of other vehicles during driving, resulting in inaccurate trajectory planning for the vehicle itself. Summary of the invention

[0004] The present application at least provides a trajectory planning method and related devices, vehicles, and storage media.

[0005] The present application provides a trajectory planning method, including: obtaining vehicle proxy information and road environment information, the vehicle proxy information carries the vehicle driving information of a target vehicle and the vehicle driving information of other vehicles, and the other vehicles are vehicles in the same driving area as the target vehicle; based on the vehicle proxy information and the road environment information, respectively determining a number of first driving trajectories of the target vehicle in a future preset time period and a number of second driving trajectories of other vehicles in a future preset time period; using a trained trajectory interaction module to process the number of first driving trajectories and the number of second driving trajectories to obtain a target predicted trajectory of the target vehicle, the trajectory interaction module being constructed based on game methodology.

[0006] The present application provides a trajectory planning device, including: an information acquisition unit, a trajectory prediction unit, and a game unit. The information acquisition unit is used to acquire vehicle proxy information and road environment information, the vehicle proxy information carries the vehicle driving information of the target vehicle and the vehicle driving information of other vehicles, and the other vehicles are vehicles in the same driving area as the target vehicle; the trajectory prediction unit is used to determine several first driving trajectories of the target vehicle in a future preset time period and several second driving trajectories of other vehicles in a future preset time period based on the vehicle proxy information and the road environment information; the game unit is used to process several first driving trajectories and several second driving trajectories using a trained trajectory interaction module to obtain a target predicted trajectory of the target vehicle, and the trajectory interaction module is constructed based on the game methodology.

[0007] The present application provides a vehicle, including a memory and a processor coupled to each other, wherein the processor is used to execute program instructions stored in the memory to implement the above-mentioned trajectory planning method.

[0008] The present application provides a computer-readable storage medium on which program instructions are stored. When the program instructions are executed by a processor, the above-mentioned trajectory planning method is implemented.

[0009] In the above scheme, by first estimating several possible driving trajectories of the target vehicle and other vehicles in a preset time period in the future according to the vehicle driving information and road environment information of the target vehicle and other vehicles, and then processing the possible driving trajectories of the target vehicle and other vehicles according to the trained trajectory interaction module, the appropriate predicted trajectory of the target vehicle can be determined. That is, this scheme fully considers the interaction between the possible driving trajectories of the target vehicle and other vehicles on the road, so as to determine the appropriate predicted trajectory of the target vehicle. Compared with first unilaterally predicting the trajectories of other vehicles and then predicting the trajectory of the target vehicle based on the possible trajectories of other vehicles, the predicted trajectory of the target vehicle determined by the present application is more accurate.

[0010] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The drawings herein are incorporated into the specification and constitute a part of the specification. These drawings illustrate embodiments consistent with the present application and are used together with the specification to illustrate the technical solution of the present application.

[0012] Figure 1 It is a flow chart of an embodiment of a trajectory planning method of the present application;

[0013] Figure 2 is a road schematic diagram shown in an embodiment of the trajectory planning method of the present application;

[0014] Figure 3 yes Figure 1 Schematic diagram of the sub-process of step S12;

[0015] Figure 4a yes Figure 3 Schematic diagram of the sub-process of step S122;

[0016] Figure 4b This is a schematic diagram showing the execution flow of the edge prediction module in an embodiment of the trajectory planning method of the present application;

[0017] Figure 5a yes Figure 1 Schematic diagram of the sub-process of step S13;

[0018] Figure 5bThis is a schematic diagram showing the execution flow of a trajectory interaction module according to an embodiment of a trajectory planning method of the present application;

[0019] Figure 6 yes Figure 5a Schematic diagram of the sub-process of step S131;

[0020] Figure 7 is a flow chart showing the first training step of an embodiment of the trajectory planning method of the present application;

[0021] Figure 8 is a flow chart showing the second training step in an embodiment of the trajectory planning method of the present application;

[0022] Figure 9a It is another flow chart of an embodiment of the trajectory planning method of the present application;

[0023] Figure 9b This is a schematic diagram showing the future motion trajectories of different vehicles in an embodiment of the trajectory planning method of the present application;

[0024] Fig.10 It is a structural schematic diagram of an embodiment of a trajectory planning device of the present application;

[0025] Fig.11 It is a schematic diagram of the framework of an embodiment of a vehicle of the present application;

[0026] Fig.12 A schematic diagram of a framework of an embodiment of a computer-readable storage medium of the present application. DETAILED DESCRIPTION

[0027] The scheme of the embodiment of the present application is described in detail below in conjunction with the drawings of the specification.

[0028] In the following description, for the purpose of explanation rather than limitation, specific details such as specific system structures, interfaces, and technologies are provided to facilitate a thorough understanding of the present application.

[0029] The term "and / or" in this article is only a description of the association relationship of associated objects, indicating that there may be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the objects associated before and after are in an "or" relationship. In addition, "many" in this article means two or more than two. In addition, the term "at least one" in this article means any combination of at least two of any one or more of a plurality of, for example, including at least one of A, B, and C, can mean including any one or more elements selected from the set consisting of A, B, and C.

[0030] The execution subject of the trajectory planning method may be a trajectory planning device, which may be any terminal device or server or other processing device capable of executing the method embodiment of the present application, wherein the terminal device may be a vehicle, an electronic device, a user equipment (UE), a mobile device, a user terminal, a terminal, a cellular phone, a cordless phone, a personal digital assistant (PDA), a handheld device, a computing device, a vehicle-mounted device, a wearable device, etc. In some possible implementations, the trajectory planning method may be implemented by a processor calling a computer-readable instruction stored in a memory.

[0031] See also Figure 1 , Figure 1 This is a flow chart of an embodiment of the trajectory planning method of the present application. Specifically, it may include the following steps:

[0032] Step S11: Acquire vehicle agent information and road environment information.

[0033] The vehicle proxy information carries the vehicle driving information of the target vehicle and the vehicle driving information of other vehicles, where the other vehicles are vehicles in the same driving area as the target vehicle. Among them, the target vehicle is a vehicle traveling on the road. Specifically in this embodiment, the target vehicle is an automatic driving or assisted driving vehicle whose driving trajectory is to be decided. The driving area here can be an area in the road network, in which other vehicles and the target vehicle may be traveling on the same road, passing the same driving position within a short time difference, or the two vehicles may meet somewhere on the road, etc. Exemplarily, the target vehicle can be a vehicle executing the trajectory planning method of the present application, such as Figure 2 As shown, the target vehicle may be vehicle 2, and the other vehicle may be vehicle 1 at the same intersection as vehicle 2. In other embodiments, the trajectory planning method of the present application may also be applied to any road, not limited to Figure 2The intersection shown. The number of other vehicles can be one or more. The vehicle driving information may include the speed, acceleration and / or driving position at several moments during the vehicle driving process. For example, the position may include the position of the vehicle at the current moment, and the position may also include the position at several historical moments. Similarly, the speed may include the speed of the vehicle at the current moment, and may also include the speed of the vehicle at several historical moments. The acceleration may include the acceleration of the vehicle at the current moment, and may also include the acceleration of the vehicle at several historical moments. The method of obtaining the vehicle proxy information may be to first obtain the vehicle driving information of the target vehicle and obtain the vehicle driving information of other vehicles, and then merge the vehicle driving information of the target vehicle with the vehicle driving information of other vehicles to obtain the vehicle proxy information. Among them, the method of obtaining the vehicle driving information of other vehicles may be to send a data acquisition request to other vehicles and then receive the vehicle driving information sent by other vehicles based on the data acquisition request. The data acquisition request may include the identification of the target vehicle and / or the vehicle driving information of the target vehicle, so as to facilitate other vehicles to plan the trajectory according to the vehicle driving information of the target vehicle. That is, the vehicle driving information between vehicles can be obtained through data sharing, and the specific data sharing method is not specifically limited here. That is, in a connected vehicle environment, vehicles can exchange data in real time with each other and with roadside infrastructure, sharing information such as their location, speed, acceleration, or planned trajectory.

[0034] The road environment information may include lane information, road indication information and / or the drivable area identification of the current road, etc. The road environment information may be obtained by pre-establishing a high-precision map, and obtaining the road environment information of the current road from the high-precision map according to the current position of the target vehicle.

[0035] Step S12: Based on the vehicle agent information and the road environment information, a plurality of first driving trajectories of the target vehicle in a future preset time period and a plurality of second driving trajectories of other vehicles in a future preset time period are respectively determined.

[0036] Based on the vehicle agent information, the possible driving trajectories of the target vehicle and other vehicles can be determined, and then based on the road environment information, the driving trajectory within the drivable area of ​​the road can be determined. That is, the road environment information can be considered to constrain the driving trajectory, so that the driving trajectory determined under the constraints of the road environment information is more in line with driving regulations, that is, the driving trajectory determined is more accurate.

[0037] Step S13: using a trained trajectory interaction module to process a plurality of first driving trajectories and a plurality of second driving trajectories for game playing, to obtain a target predicted trajectory of the target vehicle.

[0038] The trajectory interaction module is a processing module constructed based on game methodology, and can be constructed and pre-trained based on a neural network model. Here, the trajectory interaction module uses a game method to process the first driving trajectory and the second driving trajectory in the input value, and finds the optimal combination of the first driving trajectory and the second driving trajectory. Optionally, the first driving trajectory in the optimal combination of the first driving trajectory and the second driving trajectory is used as the target predicted trajectory of the target vehicle.

[0039] Exemplarily, each first driving trajectory is combined with each second driving trajectory to determine the suitability of the combination of each first driving trajectory and each second driving trajectory, such as whether the two driving trajectories overlap or whether there is a risk of a traffic accident between the two driving trajectories.

[0040] In the above scheme, by first estimating several possible driving trajectories of the target vehicle and other vehicles in a preset time period in the future according to the vehicle driving information and road environment information of the target vehicle and other vehicles, and then determining a suitable predicted trajectory of the target vehicle by gambling with the possible driving trajectories of the target vehicle and other vehicles. That is, this scheme fully considers the interaction between the possible driving trajectories of the target vehicle and other vehicles on the road, so as to determine a suitable predicted trajectory of the target vehicle. Compared with first unilaterally predicting the trajectories of other vehicles and then predicting the trajectory of the target vehicle according to the possible trajectories of other vehicles, the predicted trajectory of the target vehicle determined by the present application is more accurate.

[0041] In some embodiments, see Figure 3 , the above step S12 may include the following steps:

[0042] Step S121: extracting features from the vehicle agent information and the road environment information respectively to obtain target vehicle features and target road features.

[0043] Exemplarily, the same encoder can be used to extract features from vehicle agent information and road environment information to obtain target vehicle features and target road features, or a vehicle encoder can be used to extract features from vehicle agent information to obtain target vehicle features, and a map encoder can be used to extract features from road environment information to obtain target road features. The vehicle encoder and the map encoder can be attention-based encoders that can convert vehicle agent information and road environment information into high-dimensional latent state vectors to obtain complex relationships between vehicle agent information and complex relationships in road environment information. The vehicle encoder and the map encoder can be components of a trajectory planning model or can be set independently of the trajectory planning model.

[0044] The road environment information can be a road map, which can be represented by R∈[NP ,N V ,F], the road map can be composed of polygonal road features, each polygon is a collection of lines and the road map has type annotations for detailed representation, such as annotations of drivable areas, annotations of lane lines, etc., where, N P Represents the total number of polygons, N V Represents the number of point vectors of each polygon, and the feature size is F. Vehicle agent information can include features such as position, speed, and acceleration at past and current time. Vehicle agent information can be expressed as A∈[N A ,N T ,F A ], where N A Represents the number of vehicle proxy information, each vehicle proxy information is the information of a vehicle, and the number of vehicle proxy information is the number of vehicles, N T Indicates the length of time, if N T Equal to 20 means 20 frames of data with a time interval of 0.1s, F A Represents the feature dimension.

[0045] Step S122: Based on the target vehicle characteristics and the target road characteristics, a plurality of candidate first driving trajectories and a plurality of candidate second driving trajectories are obtained through a pre-trained edge prediction module.

[0046] In some embodiments, the trajectory planning method provided by the present application can be executed by a trajectory planning model, and the above-mentioned trajectory interaction module and edge prediction module can be modules in the trajectory planning model. The above-mentioned step S122 can be executed by the edge prediction module in the trajectory planning model, please refer to Figure 4a and Figure 4b , Figure 4a yes Figure 3 The sub-process diagram of step S122 in FIG. Figure 4b This is a schematic diagram showing the execution flow of the edge prediction module in an embodiment of the trajectory planning method of the present application.

[0047] The edge prediction module includes a second scene cross attention module. The second scene cross attention module may be a pre-trained cross attention module. Figure 4a The above step S122 may specifically include the following steps:

[0048] Step S1221: Determine the initial intended trajectories of the target vehicle and other vehicles within a future preset time period based on the third motion query data and the fourth motion query data.

[0049] Among them, in this scheme, based on the target vehicle characteristics and the target road characteristics, obtaining multiple candidate first driving trajectories and multiple candidate second driving trajectories can be considered as a decoding process, and a decoder model similar to DETR (DEtectionTRansformer) can be used in this scheme. The third motion query data and the fourth motion query data are obtained by pre-training, and the third motion query data carries the first intention trajectory endpoint information of the target vehicle in the future preset time period, and the fourth motion query data carries the second intention trajectory endpoint of other vehicles in the future preset time period. In some embodiments, the third motion query data and the fourth motion query data can also carry the intended position of at least one intermediate time point in the future preset time period. When the third motion query data and the fourth motion query data are trained, a set of initial third motion query data and initial fourth motion query data can be generated first, and then continuously iterated and updated in the training process of the following edge prediction module. When the difference between the predicted trajectory output by the edge prediction module and the true value trajectory meets the stop training condition, the third motion query data and the fourth motion query data used in the last iteration are used as the final third motion query data and the final fourth motion query data. In some embodiments, the third motion query data and the fourth motion query data in different traffic scenarios may be different. For example, in the crossroad scenario, the trajectory of each vehicle may be straight, left turn, right turn or U-turn, but in the central roundabout scenario, the trajectory of the vehicle may be entering the roundabout, continuing along the roundabout or exiting the roundabout. The different intentions of the vehicles in different scenarios will lead to differences in the trajectories of the vehicles. Therefore, this solution can set different third motion query data and fourth motion query data for different traffic scenarios. If there are multiple third motion query data and fourth motion query data, the third motion query data and fourth motion query data that best match the road environment information can be selected based on the input road environment information. Or in other embodiments, the third motion query data and the fourth motion query data carry the intended trajectory endpoint information in a variety of traffic scenarios. By combining with the road environment information, the weight of each intended trajectory endpoint information can be determined, and the trajectory intention of each vehicle can be determined based on the weight. In other embodiments, because there are differences in the drivable areas in different traffic scenarios or on different roads, the endpoint information of each intended trajectory in the third motion query data and the fourth motion query data can be preprocessed according to the road environment information to remove the endpoint information of the intended trajectory outside the drivable area, so that the planned trajectory is more accurate.

[0050] Optionally, the above-mentioned edge prediction module includes a second self-attention module, which can be a trained module for predicting vehicle trajectories. The above-mentioned method of determining the initial intention trajectory of the target vehicle and other vehicles in a preset time period in the future based on the third motion query data and the fourth motion query data can be: based on the preset vehicle characteristics of the target vehicle and the third motion query data, determine the value data, query data and index data of the fifth key-value pair. The fifth key-value pair is used as the input of the second self-attention module to obtain the initial intention trajectory of the target vehicle. For example, the preset vehicle characteristics are fused with the third motion query data to generate the value data, query data and index data in the fifth key-value pair. The method of generating value data, index data and query data in a key-value pair based on one data in the DETR decoder model is relatively conventional and will not be repeated here. Among them, the preset vehicle characteristics are also obtained by pre-training, and the preset vehicle characteristics can be a high-dimensional hidden state vector. Similarly, the initial intention trajectory of other vehicles in a preset time period in the future can be determined based on the preset vehicle characteristics of other vehicles and the fourth motion query data, together with the value data, query data and index data of other key-value pairs, and then the other key-value pairs are used as inputs of other second self-attention modules to obtain the initial intention trajectory of other vehicles.

[0051] In some embodiments, the edge prediction module includes multiple cascaded attention modules, each of which includes a second self-attention module, a second motion cross-attention module, and a second scene cross-attention module. The output result of the previous second scene cross-attention module is used as the preset vehicle feature in the second self-attention module in the next attention module, and the output result of the second scene cross-attention module in the last attention module is input into the trajectory scoring head and the first prediction head to obtain the scores of each candidate first driving trajectory, each candidate second driving trajectory, and each candidate first driving trajectory. Among them, the preset vehicle feature in the second self-attention module in the multiple cascaded attention modules is 0 or other preset values ​​obtained after the edge prediction module training is completed.

[0052] Step S1222: Determine the query data in the fourth key-value pair using the initial intention trajectory of the target vehicle and the initial intention trajectory of other vehicles; and determine the value data and index data in the fourth key-value pair using the target vehicle characteristics and the target road characteristics.

[0053] When modeling the problem as a symmetric game problem, it is crucial that the model maintains an equal representation of each traffic participant in the scene regardless of the central vehicle transformation or coordinate rotation offset. Therefore, in this solution, relative position embedding is used to distinguish the local coordinate frames of each vehicle, allowing symmetric encoding through potential modeling. In some embodiments, the edge prediction module includes a second motion cross-attention module, which is used to extract features. Optionally, the above-mentioned first intention trajectory endpoint information includes the position and / or direction angle of the first intention trajectory endpoint, and the second intention trajectory endpoint information includes the position and / or direction angle of the second intention trajectory endpoint. On this basis, the above-mentioned method of determining the query data in the fourth key-value pair using the initial intention trajectory of the target vehicle and the initial intention trajectory of other vehicles can be: determining the initial value data and initial index data in the sixth key-value pair based on the initial intention trajectory of other vehicles. Obtain the relative position information between the first intention trajectory endpoint and the second intention trajectory endpoint. Then, the relative position information is fused to the initial value data and initial index data in the sixth key-value pair to obtain the value data and index data in the sixth key-value pair. The sixth key-value pair is then used as the input of the second motion cross-attention module to obtain the query data in the fourth key-value pair.

[0054] Please continue reading Figure 4b , in this scheme, the third motion query data is embedded in the query position, I∈[N A ,N M ,3],N M represents the number of future modes (i.e., the number of trajectory endpoints or the number of trajectories, e.g., N M =64), N A represents the number of vehicle agent information, 3 is the feature dimension. The position and direction are represented by I P , I θ In the training process, the initial value of I can be obtained by clustering the positions in the vehicle agent information used in the training process, and I is continuously updated and iterated during the training process. For example, the intended trajectory endpoints in each first trajectory and each second trajectory obtained by the edge prediction module during the training process can be clustered, and then I can be updated according to the clustering result.

[0055] In some embodiments, the third motion query data and the fourth motion query data are firstly respectively encoded with absolute positions to obtain the absolute position of the third motion query data and the absolute position of the fourth motion query data. For example, for the first vehicle agent information, the absolute position encoding method may be: i =MLP(PE(I i )),I i represents the intended end point of the ith trajectory, e iRepresents the corresponding absolute position code, i and j∈NM. Similarly, e j In order to better learn the variability of the modality in the self-attention layer, the paired motion position embedding is introduced, and the relative position encoding e ij The method is as follows:

[0056]

[0057] Among them, e ij Represents the relative position encoding between the end point of the i-th intention trajectory and the end point of the j-th intention trajectory. represents the position of the end point of the j-th intention trajectory, Indicates the direction of the end point of the jth intention trajectory. In formula (1), the Fourier position embedding is used to model the relative coordinate relationship. The initial motion i∈N M and j∈N M The relative position, direction and distance between them constitute the components.

[0058] Based on this, the output of the second motion cross attention module It can be:

[0059]

[0060] Among them, q(e i +m i ),k(e ij +e j +m j ),v(e j +m j ) are respectively the query data, index data and value data of the sixth key-value pair, m j The result of the second scene cross attention module output about the j-th intention trajectory endpoint in the previous attention module, that is, the corresponding preset vehicle feature. i Similarly, the second scene cross attention module in the previous attention module outputs the result about the end point of the i-th intention trajectory, which is the corresponding preset vehicle feature. j +m j represents the initial intention trajectory obtained based on the absolute position encoding of the end point of the j-th intention trajectory and the corresponding preset vehicle features, e i +m i represents the initial intention trajectory obtained based on the absolute position encoding of the end point of the i-th intention trajectory and the corresponding preset vehicle features.

[0061] In some embodiments, Figure 4bAs shown, the value data and index data in the fourth key-value pair may also be embedded in relative positions according to the above scheme. In other embodiments, the value data and index data in the fourth key-value pair may not be embedded in relative positions. The specific implementation method is not specifically limited here.

[0062] Step S1223: Use the fourth key-value pair as input to the second scene cross-attention module to obtain multiple candidate first driving trajectories and multiple candidate second driving trajectories.

[0063] In some embodiments, the output features of the second scene cross-attention module can be input into the first prediction head to obtain multiple candidate first driving trajectories and multiple candidate second driving trajectories. In other embodiments, the output features of the second scene cross-attention module can also be input into the trajectory scoring head to obtain the scores of each candidate first driving trajectory and each candidate second driving trajectory.

[0064] Optionally, the second scene cross attention module is constructed using a standard mutual inductor cross attention (MHCA) layer and a multi-head feedforward network (FFN) layer. To avoid redundancy, this article omits the elaboration of this part. Subsequently, a Gaussian mixture model is used to construct the trajectory of each subject. In addition, the score of each mode is predicted. That is, after the first prediction head and the trajectory scoring head obtain the candidate first trajectories of each vehicle and the corresponding scores, these scores are used to decode the Gaussian mixture model (GMM) components and scores of the vehicle, respectively, to obtain the final scores of each candidate first driving trajectory and each candidate second driving trajectory.

[0065] Step S123: using at least part of the plurality of candidate first driving trajectories as a plurality of first driving trajectories, and using at least part of the plurality of candidate second driving trajectories as a plurality of second driving trajectories.

[0066] Exemplarily, several candidate first driving trajectories can be selected from high to low according to the scores of each candidate first driving trajectory as several first driving trajectories, and several candidate second driving trajectories can be selected from high to low according to the scores of each candidate second driving trajectory as several second driving trajectories. In some other embodiments, the corresponding first driving trajectory and second driving trajectory can be selected using non-maximum suppression. In some application scenarios, the output of the edge prediction module has a total of 64 predicted trajectories. In order to reduce this complexity, non-maximum suppression (NMS) is used to select the top 6 candidate first trajectories and candidate second driving trajectories for each vehicle agent information, and each vehicle agent information corresponds to a vehicle. Subsequently, the future predictions of the interacting vehicle pairs are combined to obtain a total of 36 potential vehicle pairs, and the 36 potential trajectory pairs can be expressed as: J∈[N A ,N A,36,2D],m ai represents the first driving trajectory of vehicle a, m bj represents the jth second driving trajectory of vehicle b, N A Represents the number of vehicles, that is, the number of vehicle agent information, 2D represents In some other embodiments, the fourth key-value pair may be input into the matching head to obtain a matching score between any two candidate first driving trajectories and candidate second driving trajectories, and then 12 pairs or another number of trajectory pairs may be selected from the 36 pairs according to the matching score to obtain a plurality of first driving trajectories and a plurality of second driving trajectories.

[0067] In some embodiments, see Figure 5a and Figure 5b , the above step S13 may include the following steps:

[0068] Step S131: using a trained trajectory interaction module to perform Nash equilibrium game on a plurality of first driving trajectories and a plurality of second driving trajectories, and obtaining a game score between each first driving trajectory and each second driving trajectory.

[0069] Among them, the structure of the trajectory interaction module is similar to that of the edge prediction module.

[0070] Optionally, the trajectory interaction module includes a first motion cross attention module, which is trained in advance. Figure 6 , the implementation of step S131 may include:

[0071] Step S1311: determining query data in a first key-value pair based on a plurality of first driving trajectories, and determining value data and index data in the first key-value pair based on a plurality of second driving trajectories.

[0072] Optionally, the trajectory interaction module includes a first self-attention module, which is obtained by pre-training and can perform feature extraction. Among them, based on a number of first driving trajectories, the method of determining the query data in the first key-value pair can be: based on a number of first driving trajectories and first motion query data, determine the value data, index data and query data in the second key-value pair, the first motion query data is obtained by pre-training and the first motion query data carries the first intended trajectory endpoint information of the target vehicle in a preset time period in the future. Then, the second key-value pair is used as the input of the first self-attention module to obtain the query data in the first key-value pair.

[0073] Among them, the first motion query data and the third motion query data may be the same or different. The second motion query data and the fourth motion query data may be the same or different. In some embodiments, because there are differences in drivable areas in different traffic scenarios or on different roads, the end point information of each intended trajectory in the first motion query data and the second motion query data can be preprocessed according to the road environment information, and the end point information of the intended trajectory outside the drivable area is removed, so that the planned trajectory is more accurate.

[0074] Among them, the above-mentioned method of determining the value data and index data in the first key-value pair based on several second driving trajectories can be: based on several second driving trajectories and second motion query data, determining the value data in the first key-value pair and the index data in the second key-value pair, the second motion query data is pre-trained and the second motion query data carries the second intended trajectory endpoint information of other vehicles within a preset time period in the future.

[0075] In order to explore the future motion interactions between vehicles, which is crucial for accurate and scenario-compliant predictions, a cross-attention mechanism for the future motions of different agents is introduced in the trajectory interaction module. Different from the motion self-attention layer, the dynamic intention points of the vehicles (denoted as Y p ) to construct the initial query. However, since the dynamic intent points are defined in their respective local coordinate systems, it is necessary to transform them into the global coordinate system, denoted as y wp This transformation is based on the global position Pa∈[NA,2] and direction angle Ha∈[NA,2] of the agent history, thus facilitating information propagation between intent queries of different agent histories. The global coordinates of vehicle a are Among them, the i-th trajectory of vehicle a is recorded as The jth trajectory of vehicle b is recorded as and The relative position between It is expressed as:

[0076]

[0077] In some embodiments, the first intended trajectory endpoint information includes the position and / or azimuth of the first intended trajectory endpoint, and the second intended trajectory endpoint information includes the position and / or azimuth of the second intended trajectory endpoint. The above-mentioned method of determining the value data in the first key-value pair and the index data in the second key-value pair based on a plurality of second driving trajectories and second motion query data may include: determining the initial value data in the first key-value pair and the initial index data in the first key-value pair based on a plurality of second driving trajectories and second motion data. Obtain the relative position information between the first intended trajectory endpoint and the second intended trajectory endpoint. Fuse the relative position information into the initial value data and the initial index data to obtain the value data and the index data in the first key-value pair.

[0078] Specifically, the output of the first motion criss-cross attention module is:

[0079]

[0080] Among them, a, b belong to N A , i,j belongs to N M Optionally, the trajectory interaction module includes multiple interactive attention modules (for example, including K) arranged in cascade, each interactive attention module includes a first self-attention module, a first motion cross-attention module and a first scene cross-attention module, the input of the non-first interactive attention module is the output of the first scene cross-attention module in the previous interactive attention module, and the output result of the first scene cross-attention module in the last interactive attention module is input into the second prediction head and the matching score head to obtain the score of each trajectory matching pair and each trajectory match. The input of the first attention module in the first interactive attention module can be obtained based on the output of the recurrent neural network (GRU) and the first motion query data, and the input of the recurrent neural network is several first driving trajectories of the target vehicle. Specifically, the first motion query data can be absolutely position encoded, and the data after absolute position encoding is combined with the output of the recurrent neural network to obtain the input of the first self-attention module. Among them, the attention modules mentioned in this scheme can all be multi-head attention modules.

[0081] Step S1312: Use the first key-value pair as input to the first motion cross-attention module to obtain game data.

[0082] The game data may be the features of several first driving trajectories and several second driving trajectories or several first driving trajectories and several second driving trajectories of candidates. Exemplarily, the game data are several first driving trajectories and several second driving trajectories of candidates, that is, the several first driving trajectories and several second driving trajectories of candidates may not be very accurate, and the several first driving trajectories and several second driving trajectories of candidates can be adjusted according to the characteristics of the target vehicle and the characteristics of the target road to obtain the final first driving trajectories and the second driving trajectories. Exemplarily, the game data are the features of several first driving trajectories and several second driving trajectories, and the features of the trajectories can be the approximate position where the vehicle will turn, change lanes or stop to reduce the probability of collision with other vehicles, so that the subsequent first scene cross-attention module can predict several first driving trajectories and several second driving trajectories based on the game data, the characteristics of the target vehicle and the characteristics of the target road.

[0083] Step S1313: Based on the game data, obtain the game scores between each first driving trajectory and each second driving trajectory.

[0084] Optionally, the trajectory interaction module includes a first scene cross-attention module, and the implementation method of the above step S1313 can be: determine the value data and index data in the third key-value pair based on the target vehicle characteristics and the target road characteristics, the target vehicle characteristics are obtained by feature extraction of vehicle agent information, and the target road characteristics are obtained by feature extraction of road environment information. Then, the game data is used as the query data in the third key-value pair. The third key-value pair is then used as the input of the first scene cross-attention module, and the game score is obtained based on the output result of the first scene cross-attention module.

[0085] Step S132: Determine the target prediction trajectory based on the game score.

[0086] The game score includes the game scores of multiple game groups, each of which includes a first driving trajectory and a second driving trajectory. There are multiple ways to implement step S13. One way is to select the first driving trajectory in the game group corresponding to the highest game score as the target predicted trajectory. Another way is to determine the score of each game group, which is related to at least one of the following: the probability of collision between the first driving trajectory and the second driving trajectory, the probability that the first driving trajectory and the second driving trajectory are in the drivable area of ​​the road, and the motion stability of the first driving trajectory. Then, the target predicted trajectory is determined by combining the scores of each game group and the game scores of each game group.

[0087] In some embodiments, see Figure 7, the method also includes a first training step of the edge prediction module:

[0088] Step S21: Obtain vehicle agent sample information and road environment sample information.

[0089] The vehicle agent sample information includes driving information of a first vehicle and driving information of a second vehicle.

[0090] Step S22: extracting features from the vehicle agent sample information and the road environment sample information to obtain sample vehicle features and sample road features.

[0091] Step S23: The edge prediction module obtains a plurality of first trajectories of the first vehicle and a plurality of second trajectories of the second vehicle based on the sample vehicle features and the sample road features.

[0092] Step S24: Determine a negative log-likelihood loss based on a first difference between at least one first trajectory and a true trajectory of the first vehicle, and based on a second difference between at least one second trajectory and a true trajectory of the second vehicle.

[0093] Optionally, for the predicted Gaussian mixture model at a specific time step in the future, use the negative log-likelihood loss L GMM To make the first driving trajectories Y of the vehicle and the real position Y gt The detailed loss formula is as follows, please refer to formula (5):

[0094] L GMM =-log(fh(Y gt -Y))-log(S) formula (5);

[0095] Among them, fh(Y gt -Y) is the positive Gaussian component selected for optimization. The positive Gaussian component is selected by identifying the intent query closest to the endpoint of the ground truth (GT) trajectory. S is the predicted probability of the selected Gaussian positive component.

[0096] Step S25: Based on the negative log-likelihood loss, adjust the parameters in the edge prediction module.

[0097] Optionally, the first training step may further include the following steps: determining a regularization loss based on a second difference between a target first trajectory among multiple first trajectories and a true trajectory of the first vehicle, and a third difference between a target second trajectory among multiple second trajectories and a true trajectory of the second vehicle, wherein the target first trajectory is the first trajectory among multiple first trajectories that has the highest match with the true trajectory of the first vehicle, and the target second trajectory is the second trajectory among multiple second trajectories that has the highest match with the true trajectory of the second vehicle. In the above formula (5), the cross entropy loss is used to maximize the probability of the selected Gaussian positive classification. On this basis, the above step S25 may include the following steps: adjusting the parameters in the edge prediction module based on the negative log-likelihood loss and the regularization loss. The negative log-likelihood loss and the regularization loss determine L, and L is used to adjust the parameters in the edge prediction module, where L=L GMM +L REG . Optionally, L REG A winner-takes-all strategy is adopted to optimize the position output, focusing backpropagation only on the best trajectory.

[0098] Optionally, multiple first trajectories and multiple second trajectories can constitute multiple trajectory matching groups, each trajectory matching group includes a first trajectory and a second trajectory. The first training step may also include the following steps: based on the first difference between the first target trajectory matching group and the true value trajectory of the first vehicle and the second vehicle, determine the binary cross entropy loss, the first target trajectory matching group is the trajectory matching group with the highest matching degree with the trajectory true value group among the trajectory matching groups, and the trajectory true value group includes the trajectory of the first vehicle and the second vehicle. On this basis, the above step S25 may include the following steps: based on the negative log-likelihood loss and the binary cross entropy loss, adjust the parameters in the edge prediction module.

[0099] Each trajectory pair is assigned a score. Then, a binary cross entropy (BCE) loss is applied, treating it as a classification problem. The pair closest to the ground truth trajectory is denoted as Y ab , is considered a positive sample, while all other combinations are considered negative samples. Binary cross entropy loss L machting The determination method of can refer to formula (6):

[0100]

[0101] represents the matching score between the trajectory i of vehicle a and the trajectory j of vehicle b, represents the matching group of trajectory i of vehicle a and trajectory j of vehicle b, represents the true value trajectory of vehicle a and vehicle b. Optionally, the first 12 groups with the highest matching scores are selected from the 36 pairs of matching trajectory groups as the input of the trajectory interaction module.

[0102] In some embodiments, the above step S13 is performed by a trajectory interaction module. Figure 8 The trajectory planning method may further include a second training step for the trajectory interaction module:

[0103] Step S31: Processing a plurality of first trajectories of the first vehicle and a plurality of second trajectories of the second vehicle through a trajectory interaction module to obtain a game score of each trajectory matching group.

[0104] Optionally, step S31 may include the following steps: adding the true value trajectory of the first vehicle to the multiple first trajectories to obtain multiple first trajectories after data enhancement. And, adding the true value trajectory of the second vehicle to the multiple second trajectories to obtain multiple second trajectories after data enhancement. The multiple first trajectories after data enhancement and the multiple second trajectories after data enhancement are processed by the trajectory interaction module to obtain the game score of each trajectory matching group.

[0105] Step S32: taking the trajectory matching group corresponding to the maximum game score as the second target trajectory matching group.

[0106] Step S33: Determine the negative log-likelihood loss based on the difference between the second target trajectory matching group and the trajectory truth group Cross Entropy Loss And the regularization loss One or more of .

[0107] in,

[0108] Among them, the regularization loss method is relatively conventional and will not be described here.

[0109] Step S34: using one or more of negative log-likelihood loss, cross entropy loss and regularization loss to adjust parameters in the trajectory interaction module.

[0110] The final loss

[0111] A lot of research has proved that the trajectory planning performance of this scheme is better.

[0112] As mentioned above, in the process of vehicle trajectory planning, structured high-precision map data (road environment information) and vehicle data after perception fusion (vehicle agent information) are used as input, and the vehicle interaction problem is modeled as a symmetric game problem. The overall implementation framework includes two key components: the edge prediction module and the trajectory interaction module. The edge prediction module outputs multiple potential motion trajectories of multiple vehicles in a given traffic scene from the perspective of a single vehicle. Subsequently, the trajectory interaction module regards the multiple outputs of the edge prediction module as part of the Nash equilibrium problem. Based on the Nash equilibrium principle, this scheme develops a unique matching loss (the above-mentioned cross entropy loss) and introduces a method for data enhancement of training data based on the true value trajectory of each vehicle for training. This comprehensive approach enables a better understanding and prediction of the complex dynamics of vehicle interactions in complex traffic environments, thereby achieving more accurate and reliable decision-making for autonomous driving vehicles. In addition, the interaction module of this method can be used as a plug-and-play module to adapt to all existing mainstream edge prediction modules.

[0113] To better understand the above scenario, please refer to Figure 9a , Figure 9a FIG. 1 is another flow chart of an embodiment of the trajectory planning method of the present application. Figure 9a As shown, after obtaining the vehicle agent information and the road environment information, the vehicle agent information is input into the vehicle encoder to obtain the target vehicle feature (the vehicle latent vector, which is also the high-dimensional latent state vector mentioned above), and the road environment information is input into the map encoder to obtain the target road feature (the lane line latent vector, which is also the high-dimensional latent state vector mentioned above). Different second self-attention modules (vehicle multi-intention interactive attention modules) process the third motion query data and the fourth motion query data to obtain the initial trajectory intentions of the target vehicle and other vehicles. Then, the cross-attention modules (the second motion cross-attention module and the second scene cross-attention module) combine the target vehicle feature and the target road feature to obtain several first driving trajectories and several second driving trajectories. In the trajectory interaction module, each first driving trajectory and each second driving trajectory are gamed to obtain each trajectory matching group and the score of each trajectory matching group, so as to obtain the target trajectory finally predicted by the target vehicle according to the score of each matching group.

[0114] The trajectory planning model in this scheme can be considered as a Transformer-based encoder-decoder pipeline, which is mainly divided into two stages: (i) an attention-based encoder is used to convert vehicle and road map data into high-dimensional latent state vectors to capture the complex relationship between the agent and road map elements; (ii) the edge prediction module uses a decoder to generate multiple future motion trajectories, and the decoder applies an attention mechanism to improve prediction accuracy. (iii) The interaction prediction module models the multiple outputs of the edge prediction module as a symmetric combinatorial game problem, combines the future plan of the autonomous driving vehicle with the predicted trajectory of the environment vehicle, and outputs the final joint planning prediction result through the decoder. The encoder processes multiple data modes including road details, agent history, and interactions, and uses an attention mechanism to fully capture information. Relative position embedding is also introduced to maintain the symmetry of the problem model. The decoder extracts high-dimensional representations through the attention mechanism, applies multi-head self-attention layers and cross-attention layers, and finally generates the Gaussian mixture model (GMM) trajectory prediction of the agent. In order to improve the accuracy of the prediction model, the Negative Log-Likelihood Loss (NLL Loss) is used, which trains the model by comparing the difference between the predicted trajectory and the true trajectory. In addition, a Winner-Takes-All strategy is used, which only considers the best prediction proposal when optimizing the position output and focuses only on this proposal in the back propagation.

[0115] The game-theory-based trajectory interaction module introduces a novel matching loss mechanism to reduce learning complexity when dealing with the future motion prediction of different agents in driving scenarios, and strengthens the connection with game theory, especially the Nash equilibrium principle. In the edge prediction module, although the cross-attention of future motion is taken into account, it becomes challenging for the model to ensure reasonable attention allocation in high-complexity scenarios due to the attention allocation involving a large number of motions. To simplify the learning process, the non-maximum suppression (NMS) technique is first used to select the top six prediction results for each agent, and then they are paired to form a total of 36 possible interaction pairs. In this framework, each motion pair is assigned a score, which is completed by the matching head. Subsequently, the binary cross entropy (BCE) loss function is used to classify the motion pair closest to the ground truth as a positive sample, while all other combinations are regarded as negative samples. In addition, in order to reflect the decision-making process in actual driving behavior, the interaction prediction module of this scheme is constructed to reflect the structure of the Nash equilibrium game. For the 36 motion pair outputs generated by the marginal prediction module, this scheme applies the Nash equilibrium principle to select the most likely interaction pair.

[0116] like Figure 9bAs shown in the figure, each vehicle has multiple possible future motion trajectories at the intersection, which are represented by dotted lines. These trajectories include not only the possible directions of travel of the vehicle, but also its dynamic properties such as speed and acceleration. In order to reduce the computational complexity during prediction, the non-maximum suppression technique is first applied to the predicted trajectory of each vehicle, and the top six predicted trajectories with the highest probability are selected. These trajectories are then paired with the corresponding trajectories of other vehicles to form multiple potential interaction pairs. Each pairing is scored by the model, and the accuracy of each pairing is evaluated by a matching loss function, where the pairing closest to the actual trajectory (such as the trajectory represented by the solid line in the figure) is considered a positive sample, and other pairs are considered negative samples. This approach enables the model to more effectively learn and predict vehicle interactions in more complex traffic scenarios, such as intersections.

[0117] In addition, the trajectory interaction module continues to use the Gaussian mixture model (GMM) loss, which is a loss function specifically designed to maximize the consistency of agent pair predictions and ground truth interactions. This proposal improves the matching loss to better reflect the strategic decisions of real-life traffic interactions. If the other vehicle chooses its best motion prediction, the target vehicle is expected to choose its current best motion. Therefore, this proposal introduces the cross-entropy loss of the trajectory interaction module to optimize the prediction of the agent pair. Finally, the framework of this proposal adopts a combination of cross-entropy loss and GMM loss, as well as other regularization losses, which together constitute the loss function of this proposal. The above is a detailed introduction to the game theory-based interaction module. This section emphasizes how the model learns and optimizes in predicting agent interactions in future traffic scenarios, and how to combine game theory principles through specific loss functions to improve the accuracy of predictions and real-world applicability.

[0118] In the above scheme, in an urban environment, an autonomous vehicle uses an interaction module based on game theory to predict the possible action paths of other vehicles and pedestrians when approaching a busy intersection. The module evaluates various potential interaction scenarios and generates a series of safe and reasonable action trajectories for the vehicle, helping the vehicle make the best decision, avoid potential collisions, and efficiently pass through the intersection.

[0119] In addition, this solution can be used to exchange real-time data between vehicles and between vehicles and roadside infrastructure in the Internet of Vehicles environment, sharing information such as their location, speed, and expected trajectory. Using the interactive module based on game theory, each vehicle can more accurately predict and understand the behavior of surrounding vehicles, so as to collaborate to complete complex driving tasks such as merging and avoiding obstacles, thereby enhancing road safety and improving traffic flow.

[0120] Moreover, through the use of symmetric games and Nash equilibrium, autonomous vehicles can make decisions that are more in line with actual traffic conditions and effectively balance aggressive and cautious driving strategies. In addition, the model takes into account multiple interaction results, reduces decision uncertainty caused by ignoring complex interaction dynamics, and improves the accuracy and safety of planning.

[0121] The above solution can be applied to autonomous driving vehicles, especially for autonomous driving strategy planning in urban traffic and complex road conditions.

[0122] Those skilled in the art will appreciate that, in the above method of specific implementation, the order in which the steps are written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of the steps should be determined by their functions and possible internal logic.

[0123] See also Fig.10 , Fig.10 It is a structural diagram of an embodiment of the trajectory planning device of the present application. The trajectory planning device 40 includes an information acquisition unit 41, a trajectory prediction unit 42 and a game unit 43. The information acquisition unit 41 is used to acquire vehicle agent information and road environment information. The vehicle agent information carries the vehicle driving information of the target vehicle and the vehicle driving information of other vehicles. The other vehicles are vehicles in the driving area with the target vehicle; the trajectory prediction unit 42 is used to determine several first driving trajectories of the target vehicle in a future preset time period and several second driving trajectories of other vehicles in a future preset time period based on the vehicle agent information and the road environment information; the game unit 43 is used to process several first driving trajectories and several second driving trajectories using a trained trajectory interaction module to obtain a target predicted trajectory of the target vehicle. The trajectory interaction module is constructed based on the game methodology.

[0124] In the above scheme, by first estimating several possible driving trajectories of the target vehicle and other vehicles in a preset time period in the future according to the vehicle driving information and road environment information of the target vehicle and other vehicles, and then processing the possible driving trajectories of the target vehicle and other vehicles according to the trained trajectory interaction module, the appropriate predicted trajectory of the target vehicle can be determined. That is, this scheme fully considers the interaction between the possible driving trajectories of the target vehicle and other vehicles on the road, so as to determine the appropriate predicted trajectory of the target vehicle. Compared with first unilaterally predicting the trajectories of other vehicles and then predicting the trajectory of the target vehicle based on the possible trajectories of other vehicles, the predicted trajectory of the target vehicle determined by the present application is more accurate.

[0125] In some embodiments, the game unit 43 uses a trained trajectory interaction module to process a plurality of first driving trajectories and a plurality of second driving trajectories to obtain a target predicted trajectory of the target vehicle. The trajectory interaction module is constructed based on game methodology, including: using a trained trajectory interaction module to perform Nash equilibrium game on a plurality of first driving trajectories and a plurality of second driving trajectories to obtain a game score between each first driving trajectory and each second driving trajectories; based on the game score, determining the target predicted trajectory.

[0126] In the above scheme, by performing Nash equilibrium game on each first driving trajectory and each second driving trajectory, a better predicted trajectory of the target vehicle can be determined.

[0127] In some embodiments, the trajectory interaction module includes a first motion cross-attention module, and the game unit 43 performs Nash equilibrium game on several first driving trajectories and several second driving trajectories to obtain the game score between each first driving trajectory and each second driving trajectory, including: determining the query data in the first key-value pair based on the several first driving trajectories, and determining the value data and index data in the first key-value pair based on the several second driving trajectories; using the first key-value pair as the input of the first motion cross-attention module to obtain the game data; based on the game data, obtaining the game score between each first driving trajectory and each second driving trajectory.

[0128] In the above scheme, query data (Q), value data (V) and index data (K) of the first motion cross-attention module are generated through each first driving trajectory and each second driving trajectory, so that the game score between each first driving trajectory and each second driving trajectory can be determined under the action of the cross-attention mechanism.

[0129] In some embodiments, the trajectory interaction module includes a first self-attention module, and the game unit 43 determines the query data in the first key-value pair based on several first driving trajectories, including: determining the value data, index data and query data in the second key-value pair based on several first driving trajectories and the first motion query data, the first motion query data is pre-trained and the first motion query data carries the first intended trajectory endpoint information of the target vehicle within a preset time period in the future; taking the second key-value pair as the input of the first self-attention module to obtain the query data in the first key-value pair; and / or, based on several second driving trajectories, determining the value data and index data in the first key-value pair, including: determining the value data in the first key-value pair and the index data in the second key-value pair based on several second driving trajectories and the second motion query data, the second motion query data is pre-trained and the second motion query data carries the second intended trajectory endpoint information of other vehicles within a preset time period in the future.

[0130] In the above scheme, the value data, index data and query data in the second key-value pair are obtained by first strengthening the data of the first driving trajectories using the first motion query data obtained through training, and then the query data in the first key-value pair obtained according to the second key-value pair is more accurate. In addition, the value data and index data in the first key-value pair are obtained by strengthening the second driving trajectories according to the second motion query data and based on the strengthened second driving trajectories.

[0131] In some embodiments, the trajectory interaction module includes a first scene cross-attention module, and the game unit 43 obtains the game score between each first driving trajectory and each second driving trajectory based on the game data, including: determining the value data and index data in the third key-value pair based on the target vehicle characteristics and the target road characteristics, the target vehicle characteristics are obtained by feature extraction of vehicle agent information, and the target road characteristics are obtained by feature extraction of road environment information; using the game data as the query data in the third key-value pair; using the third key-value pair as the input of the first scene cross-attention module, and obtaining the game score based on the output result of the first scene cross-attention module.

[0132] In the above scheme, by participating in the game between several first driving trajectories and several second driving trajectories based on the combination of target vehicle characteristics and target road characteristics, the game process can refer to the road environment information and the driving information of each vehicle, so that the obtained game score is more accurate.

[0133] In some embodiments, the first intention trajectory endpoint information includes the position and / or direction angle of the first intention trajectory endpoint, and the second intention trajectory endpoint information includes the position and / or direction angle of the second intention trajectory endpoint. The game unit 43 determines the value data in the first key-value pair and the index data in the second key-value pair based on the plurality of second driving trajectories and the second motion query data, including: determining the initial value data in the first key-value pair and the initial index data in the first key-value pair based on the plurality of second driving trajectories and the second motion data; obtaining the relative position information between the first intention trajectory endpoint and the second intention trajectory endpoint; and fusing the relative position information into the initial value data and the initial index data to obtain the value data and the index data in the first key-value pair.

[0134] In the above scheme, data symmetric encoding is achieved by incorporating the relative position information between the end point of the first intention trajectory and the end point of the second intention trajectory into the corresponding value data and index data.

[0135] In some embodiments, the trajectory prediction unit 42 determines several first driving trajectories of the target vehicle in a future preset time period and several second driving trajectories of other vehicles in a future preset time period based on the vehicle agent information and the road environment information, including: extracting features of the vehicle agent information and the road environment information respectively to obtain target vehicle features and target road features; based on the target vehicle features and the target road features, predicting multiple candidate first driving trajectories and multiple candidate second driving trajectories through a pre-trained edge prediction module; using at least some of the multiple candidate first driving trajectories as several first driving trajectories, and using at least some of the multiple candidate second driving trajectories as several second driving trajectories.

[0136] In the above scheme, a plurality of candidate first driving trajectories and candidate second driving trajectories are obtained by first using the target vehicle characteristics and the target road characteristics, and then some candidate first driving trajectories and candidate second driving trajectories are screened from the plurality of candidate first driving trajectories and candidate second driving trajectories as the final plurality of first driving trajectories and plurality of second driving trajectories.

[0137] In some embodiments, the edge prediction module includes a second scene cross-attention module, and the trajectory prediction unit 42 obtains multiple candidate first driving trajectories and multiple candidate second driving trajectories based on the target vehicle characteristics and the target road characteristics, including: based on the third motion query data and the fourth motion query data, determining the initial intention trajectory of the target vehicle and other vehicles in a future preset time period, the first query data and the second motion query data are pre-trained, and the third motion query data carries the first intention trajectory endpoint information of the target vehicle in the future preset time period, and the fourth motion query data carries the second intention trajectory endpoint of other vehicles in the future preset time period; using the initial intention trajectory of the target vehicle and the initial intention trajectory of other vehicles to determine the query data in the fourth key-value pair; and, using the target vehicle characteristics and the target road characteristics, determining the value data and index data in the fourth key-value pair; using the fourth key-value pair as the input of the second scene cross-attention module to obtain multiple candidate first driving trajectories and multiple candidate second driving trajectories.

[0138] In the above scheme, by first determining the initial intention trajectories of the target vehicle and other vehicles based on the third motion query data and the fourth motion query data obtained in advance through training, and then determining the input data of the second scene cross-attention module based on the target vehicle characteristics and the target road characteristics and the initial intention trajectories of each vehicle, the above-mentioned multiple candidate first driving trajectories and multiple candidate second driving trajectories can be determined, and by referring to the initial intention trajectory of each vehicle in trajectory prediction, the determined candidate driving trajectory is closer to the actual driving trajectory.

[0139] In some embodiments, the edge prediction module includes a second self-attention module, and the trajectory prediction unit 42 determines the initial intention trajectory of the target vehicle and other vehicles within a future preset time period based on the third motion query data and the fourth motion query data, including: determining the value data, query data and index data of the fifth key-value pair based on the preset vehicle characteristics of the target vehicle and the third motion query data; using the fifth key-value pair as the input of the second self-attention module to obtain the initial intention trajectory of the target vehicle.

[0140] In the above scheme, by pre-training the preset vehicle features of the target vehicle and the third motion query data, it is possible to achieve the effect of pre-estimating the initial intended trajectory of the target vehicle.

[0141] In some embodiments, the edge prediction module includes a second motion cross-attention module, the first intention trajectory endpoint information includes the position and / or direction angle of the first intention trajectory endpoint, the second intention trajectory endpoint information includes the position and / or direction angle of the second intention trajectory endpoint, and the trajectory prediction unit 42 uses the initial intention trajectory of the target vehicle and the initial intention trajectory of other vehicles to determine the query data in the fourth key-value pair, including: determining the initial value data and initial index data in the sixth key-value pair based on the initial intention trajectory of other vehicles; obtaining the relative position information between the first intention trajectory endpoint and the second intention trajectory endpoint; fusing the relative position information into the initial value data and initial index data in the sixth key-value pair to obtain the value data and index data in the sixth key-value pair; using the sixth key-value pair as the input of the second motion cross-attention module to obtain the query data in the fourth key-value pair.

[0142] In the above scheme, by referring to the relative position relationship between the end points of the two intention trajectories, the output result of the second motion cross-attention module can be determined to be more accurate.

[0143] In some embodiments, the game score includes the game scores of multiple game groups, each game group includes a first driving trajectory and a second driving trajectory, and the game unit 43 determines the target predicted trajectory based on the game scores, including: selecting the first driving trajectory in the game group corresponding to the highest game score as the target predicted trajectory; or, the game unit 43 determines the target predicted trajectory based on the game scores, including: for each game group, determining the score of the game group, the score is related to at least one of the following: the probability of collision between the first driving trajectory and the second driving trajectory, the probability that the first driving trajectory and the second driving trajectory are in a drivable area of ​​the road, and the motion smoothness of the first driving trajectory; and, combining the scores of each game group and the game scores of each game group, determining the target predicted trajectory.

[0144] In the above scheme, directly taking the first driving trajectory in the game group corresponding to the highest game score as the target prediction trajectory can simplify the trajectory planning process. In addition, by determining the collision probability between two driving trajectories in each game group, the motion stability of each driving trajectory, and whether it is always in the drivable area of ​​the road, the rationality of each driving trajectory can be determined, so that the final trajectory is more accurate.

[0145] In some embodiments, the trajectory planning device also includes a training unit (not shown) for performing a first training step for the edge prediction module: obtaining vehicle agent sample information and road environment sample information, the vehicle agent sample information including driving information of the first vehicle and driving information of the second vehicle; performing feature extraction on the vehicle agent sample information and the road environment sample information to obtain sample vehicle features and sample road features; the edge prediction module obtains multiple first trajectories of the first vehicle and multiple second trajectories of the second vehicle based on the sample vehicle features and the sample road features; determining a negative log-likelihood loss based on a first difference between at least one first trajectory and the true trajectory of the first vehicle, and based on a second difference between at least one second trajectory and the true trajectory of the second vehicle; and adjusting parameters in the edge prediction module based on the negative log-likelihood loss.

[0146] In the above scheme, by determining the negative log-likelihood loss in advance according to the first difference and the second difference, the parameters in the edge prediction module are adjusted, so that the edge prediction module can predict the trajectory of each vehicle more accurately after training.

[0147] In some embodiments, the training unit is further used to: determine the regularization loss based on a second difference between a target first trajectory among multiple first trajectories and a true trajectory of the first vehicle, and a third difference between a target second trajectory among multiple second trajectories and the true trajectory of the second vehicle, the target first trajectory being the first trajectory among the multiple first trajectories with the highest match to the true trajectory of the first vehicle, and the target second trajectory being the second trajectory among the multiple second trajectories with the highest match to the true trajectory of the second vehicle; and adjust the parameters in the edge prediction module based on the negative log-likelihood loss, including: adjusting the parameters in the edge prediction module based on the negative log-likelihood loss and the regularization loss.

[0148] In the above scheme, the first trajectory with the highest matching degree with the true trajectory of the first vehicle among the multiple first trajectories is taken as the positive sample, and the other data are taken as

[0149] In some embodiments, multiple first trajectories and multiple second trajectories can constitute multiple trajectory matching groups, each trajectory matching group includes a first trajectory and a second trajectory, and the training unit is also used to: determine the binary cross entropy loss based on the first difference between the first target trajectory matching group and the true value trajectory of the first vehicle and the second vehicle, the first target trajectory matching group is the trajectory matching group with the highest matching degree with the trajectory true value group among the trajectory matching groups, and the trajectory true value group includes the trajectory of the first vehicle and the second vehicle; the training unit adjusts the parameters in the edge prediction module based on the negative log-likelihood loss, including: adjusting the parameters in the edge prediction module based on the negative log-likelihood loss and the binary cross entropy loss.

[0150] In the above scheme, the first trajectory with the highest matching degree with the true trajectory of the first vehicle among multiple first trajectories is taken as a positive sample, and the other first trajectories are taken as negative samples, and the second trajectory with the highest matching degree with the true trajectory of the second vehicle among multiple second trajectories is taken as a positive sample, and the other second trajectories are taken as negative samples to determine the binary cross entropy loss, and combined with the negative log-likelihood loss, the adjusted edge prediction module can more accurately predict the trajectory of each vehicle.

[0151] In some embodiments, the training unit is further used to perform a second training step for the trajectory interaction module: processing multiple first trajectories of the first vehicle and multiple second trajectories of the second vehicle through the trajectory interaction module to obtain the game scores of each trajectory matching group respectively; taking the trajectory matching group corresponding to the maximum game score as the second target trajectory matching group; based on the difference between the second target trajectory matching group and the trajectory true value group, determining one or more of the negative log-likelihood loss, the cross entropy loss and the regularization loss; using one or more of the negative log-likelihood loss, the cross entropy loss and the regularization loss to adjust the parameters in the trajectory interaction module.

[0152] In the above scheme, at least one loss is determined by utilizing the difference between the second target trajectory matching group and the trajectory truth group, so that the trajectory interaction module after parameter adjustment can better determine the game interaction score between each trajectory.

[0153] In some embodiments, a plurality of first trajectories of a first vehicle and a plurality of second trajectories of a second vehicle are processed by a trajectory interaction module to obtain a gaming score of each trajectory matching group, respectively, including: adding the true value trajectory of the first vehicle to the plurality of first trajectories to obtain a plurality of first trajectories after data enhancement; and adding the true value trajectory of the second vehicle to the plurality of second trajectories to obtain a plurality of second trajectories after data enhancement; and processing the plurality of first trajectories after data enhancement and the plurality of second trajectories after data enhancement by the trajectory interaction module to obtain a gaming score of each trajectory matching group.

[0154] In the above scheme, during the training process, data enhancement is performed on multiple first trajectories and multiple second trajectories by adding the true value trajectories of the first vehicle and the second vehicle, so that the trained trajectory interaction module can better determine the game interaction scores between the trajectories.

[0155] See also Fig.11 , Fig.11 The vehicle 50 includes a memory 51 and a processor 52 coupled to each other, and the processor 52 is used to execute program instructions stored in the memory 51 to implement the steps in any of the above trajectory planning method embodiments.

[0156] Specifically, the processor 52 is used to control itself and the memory 51 to implement the steps of any of the above trajectory planning method embodiments, or to implement the steps in any of the above trajectory planning method embodiments. The processor 52 can also be referred to as a CPU (Central Processing Unit). The processor 52 may be an integrated circuit chip with signal processing capabilities. The processor 52 may also be a general-purpose processor, a digital signal processor (Digital Signal Processor, DSP), an application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), a field programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. In addition, the processor 52 may be implemented by an integrated circuit chip.

[0157] In the above scheme, by first estimating several possible driving trajectories of the target vehicle and other vehicles in a preset time period in the future according to the vehicle driving information and road environment information of the target vehicle and other vehicles, and then processing the possible driving trajectories of the target vehicle and other vehicles according to the trained trajectory interaction module, the appropriate predicted trajectory of the target vehicle can be determined. That is, this scheme fully considers the interaction between the possible driving trajectories of the target vehicle and other vehicles on the road, so as to determine the appropriate predicted trajectory of the target vehicle. Compared with first unilaterally predicting the trajectories of other vehicles and then predicting the trajectory of the target vehicle based on the possible trajectories of other vehicles, the predicted trajectory of the target vehicle determined by the present application is more accurate.

[0158] See also Fig.12 , Fig.12The computer-readable storage medium 60 stores program instructions 601 that can be executed by a processor, and the program instructions 601 are used to implement the steps in any of the above trajectory planning method embodiments.

[0159] In the above scheme, by first estimating several possible driving trajectories of the target vehicle and other vehicles in a preset time period in the future according to the vehicle driving information and road environment information of the target vehicle and other vehicles, and then processing the possible driving trajectories of the target vehicle and other vehicles according to the trained trajectory interaction module, the appropriate predicted trajectory of the target vehicle can be determined. That is, this scheme fully considers the interaction between the possible driving trajectories of the target vehicle and other vehicles on the road, so as to determine the appropriate predicted trajectory of the target vehicle. Compared with first unilaterally predicting the trajectories of other vehicles and then predicting the trajectory of the target vehicle based on the possible trajectories of other vehicles, the predicted trajectory of the target vehicle determined by the present application is more accurate.

[0160] In some embodiments, the functions or modules included in the device provided by the embodiments of the present disclosure can be used to execute the method described in the above method embodiments. The specific implementation can refer to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0161] The above description of various embodiments tends to emphasize the differences between the various embodiments. The same or similar aspects can be referenced to each other, and for the sake of brevity, they will not be repeated herein.

[0162] In the several embodiments provided in the present application, it should be understood that the disclosed methods and devices can be implemented in other ways. For example, the device implementation described above is only schematic. For example, the division of modules or units is only a logical function division. There may be other division methods in actual implementation, such as units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of devices or units can be electrical, mechanical or other forms.

[0163] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0164] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) or a processor (processor) to perform all or part of the steps of each implementation method of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program code.

[0165] If the technical solution of this application involves personal information, the product using the technical solution of this application has clearly informed the personal information processing rules and obtained the individual's voluntary consent before processing the personal information. If the technical solution of this application involves sensitive personal information, the product using the technical solution of this application has obtained the individual's separate consent before processing the sensitive personal information, and at the same time meets the "explicit consent" requirement. For example, on personal information collection devices such as cameras, clear and prominent signs are set to inform that the personal information collection scope has been entered and personal information will be collected. If the individual voluntarily enters the collection scope, it is deemed that he or she agrees to the collection of his or her personal information; or on the device that processes personal information, the personal information processing rules are notified by obvious signs / information, and the individual's authorization is obtained through pop-up information or by asking the individual to upload his or her personal information; among them, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the type of personal information processed.

Claims

1. A trajectory planning method, characterized in that: include: Acquire vehicle proxy information and road environment information, wherein the vehicle proxy information carries vehicle driving information of the target vehicle and vehicle driving information of other vehicles, wherein the other vehicles are vehicles in the same driving area as the target vehicle; Based on the vehicle agent information and the road environment information, respectively determine a plurality of first driving trajectories of the target vehicle within a future preset time period and a plurality of second driving trajectories of the other vehicles within a future preset time period; A trained trajectory interaction module is used to process the plurality of first driving trajectories and the plurality of second driving trajectories to obtain a target predicted trajectory of the target vehicle, wherein the trajectory interaction module is constructed based on game methodology.

2. The method according to claim 1, characterized in that The method of using a trained trajectory interaction module to process the plurality of first driving trajectories and the plurality of second driving trajectories to obtain a target predicted trajectory of the target vehicle includes: Using the trained trajectory interaction module to perform Nash equilibrium game on the plurality of first driving trajectories and the plurality of second driving trajectories, to obtain a game score between each of the first driving trajectories and each of the second driving trajectories; Based on the game score, the target prediction trajectory is determined.

3. The method according to claim 2, characterized in that The gaming scores include gaming scores of a plurality of gaming groups, each of the gaming groups includes a first driving track and a second driving track, The step of determining the target prediction trajectory based on the game score includes: Select the first driving trajectory in the game group corresponding to the highest game score as the target predicted trajectory; or, The step of determining the target prediction trajectory based on the game score includes: For each of the game groups, determining a score of the game group, the score being related to at least one of: a probability of collision between a first driving trajectory and a second driving trajectory, a probability that the first driving trajectory and the second driving trajectory are in a drivable area of ​​the road, and a motion smoothness of the first driving trajectory; and The target prediction trajectory is determined by combining the scores of each of the gaming groups and the gaming scores of each of the gaming groups.

4. The method according to claim 2 or 3, characterized in that: The trajectory interaction module includes a first motion cross attention module, and the Nash equilibrium game is performed on the plurality of first driving trajectories and the plurality of second driving trajectories to obtain a game score between each of the first driving trajectories and each of the second driving trajectories, including: Determine query data in a first key-value pair based on the plurality of first driving trajectories, and determine value data and index data in the first key-value pair based on the plurality of second driving trajectories; Using the first key-value pair as input of a first motion cross-attention module to obtain game data; Based on the game data, a game score between each of the first driving trajectories and each of the second driving trajectories is obtained.

5. The method according to claim 4, characterized in that The trajectory interaction module includes a first self-attention module, and the determining of query data in a first key-value pair based on the plurality of first driving trajectories includes: Determine the value data, index data and query data in the second key-value pair based on the plurality of first driving trajectories and the first motion query data, wherein the first motion query data is pre-trained and carries the first intended trajectory endpoint information of the target vehicle within a preset time period in the future; Using the second key-value pair as input of the first self-attention module to obtain query data in the first key-value pair; And / or, determining the value data and index data in the first key-value pair based on the plurality of second driving trajectories includes: Based on the several second driving trajectories and the second motion query data, the value data in the first key-value pair and the index data in the second key-value pair are determined, wherein the second motion query data is pre-trained and carries the second intended trajectory endpoint information of the other vehicles within a preset time period in the future.

6. The method according to claim 4 or 5, characterized in that: The trajectory interaction module includes a first scene cross attention module, and obtaining a game score between each of the first driving trajectories and each of the second driving trajectories based on the game data includes: Determine value data and index data in a third key-value pair based on a target vehicle feature and a target road feature, wherein the target vehicle feature is obtained by extracting features from the vehicle proxy information, and the target road feature is obtained by extracting features from the road environment information; Using the gaming data as query data in the third key-value pair; The third key-value pair is used as the input of the first scene cross-attention module, and the game score is obtained based on the output result of the first scene cross-attention module.

7. The method according to claim 5, characterized in that The first intended trajectory endpoint information includes the position and / or direction angle of the first intended trajectory endpoint, the second intended trajectory endpoint information includes the position and / or direction angle of the second intended trajectory endpoint, and the determining, based on the plurality of second driving trajectories and the second motion query data, the value data in the first key-value pair and the index data in the second key-value pair includes: Determine the initial value data in the first key-value pair and the initial index data in the first key-value pair based on the plurality of second driving trajectories and the second motion data; Acquire relative position information between the end point of the first intention trajectory and the end point of the second intention trajectory; The relative position information is merged into the initial value data and the initial index data to obtain the value data and the index data in the first key-value pair.

8. The method according to any one of claims 1 to 7, characterized in that The determining, based on the vehicle agent information and the road environment information, a plurality of first driving trajectories of the target vehicle in a future preset time period and a plurality of second driving trajectories of the other vehicles in a future preset time period respectively, comprises: Extracting features from the vehicle agent information and the road environment information respectively to obtain target vehicle features and target road features; Based on the target vehicle characteristics and the target road characteristics, a plurality of candidate first driving trajectories and a plurality of candidate second driving trajectories are predicted by a pre-trained edge prediction module; At least part of the plurality of candidate first driving trajectories is used as the plurality of first driving trajectories, and at least part of the plurality of candidate second driving trajectories is used as the plurality of second driving trajectories.

9. The method according to claim 8, characterized in that The edge prediction module includes a second scene cross attention module, which obtains a plurality of candidate first driving trajectories and a plurality of candidate second driving trajectories based on the target vehicle feature and the target road feature, including: Based on the third motion query data and the fourth motion query data, determining the initial intended trajectories of the target vehicle and the other vehicles within the future preset time period, the third motion query data and the fourth motion query data are obtained by pre-training, and the third motion query data carries the first intended trajectory endpoint information of the target vehicle within the future preset time period, and the fourth motion query data carries the second intended trajectory endpoint information of the other vehicles within the future preset time period; Determine query data in a fourth key-value pair using the initial intention trajectory of the target vehicle and the initial intention trajectory of the other vehicles; and determine value data and index data in the fourth key-value pair using the target vehicle characteristics and the target road characteristics; The fourth key-value pair is used as the input of the second scene cross-attention module to obtain the multiple candidate first driving trajectories and the multiple candidate second driving trajectories.

10. The method according to claim 9, characterized in that The edge prediction module includes a second self-attention module, and the determining of the initial intention trajectories of the target vehicle and the other vehicles within the future preset time period based on the third motion query data and the fourth motion query data includes: Determining value data, query data, and index data of a fifth key-value pair based on the preset vehicle characteristics of the target vehicle and the third motion query data; The fifth key-value pair is used as an input of the second self-attention module to obtain an initial intention trajectory of the target vehicle.

11. The method according to claim 9, characterized in that The edge prediction module includes a second motion cross attention module, the first intention trajectory endpoint information includes the position and / or direction angle of the first intention trajectory endpoint, the second intention trajectory endpoint information includes the position and / or direction angle of the second intention trajectory endpoint, and the query data in the fourth key-value pair is determined by using the initial intention trajectory of the target vehicle and the initial intention trajectory of the other vehicles, including: Determining initial value data and initial index data in a sixth key-value pair based on the initial intention trajectory of the other vehicle; Acquire relative position information between the end point of the first intention trajectory and the end point of the second intention trajectory; Merging the relative position information into the initial value data and the initial index data in the sixth key-value pair to obtain the value data and the index data in the sixth key-value pair; The sixth key-value pair is used as the input of the second motion cross-attention module to obtain the query data in the fourth key-value pair.

12. The method according to any one of claims 8 to 11, characterized in that The method further comprises a first training step of the edge prediction module: Acquire vehicle agent sample information and road environment sample information, wherein the vehicle agent sample information includes driving information of a first vehicle and driving information of a second vehicle; Extracting features from the vehicle agent sample information and the road environment sample information to obtain sample vehicle features and sample road features; The edge prediction module obtains a plurality of first trajectories of the first vehicle and a plurality of second trajectories of the second vehicle based on the sample vehicle features and the sample road features; determining a negative log-likelihood loss based on a first difference between at least one of the first trajectories and a ground-truth trajectory of the first vehicle, and based on a second difference between at least one of the second trajectories and a ground-truth trajectory of the second vehicle; Based on the negative log-likelihood loss, parameters in the edge prediction module are adjusted.

13. The method according to claim 12, characterized in that The method further comprises: determining a regularization loss based on a second difference between a target first trajectory among the plurality of first trajectories and a true trajectory of the first vehicle, and a third difference between a target second trajectory among the plurality of second trajectories and the true trajectory of the second vehicle, wherein the target first trajectory is a first trajectory among the plurality of first trajectories with a highest degree of matching with the true trajectory of the first vehicle, and the target second trajectory is a second trajectory among the plurality of second trajectories with a highest degree of matching with the true trajectory of the second vehicle; The adjusting the parameters in the edge prediction module based on the negative log-likelihood loss includes: Based on the negative log-likelihood loss and the regularization loss, parameters in the edge prediction module are adjusted.

14. The method according to claim 12 or 13, characterized in that The plurality of first trajectories and the plurality of second trajectories can constitute a plurality of trajectory matching groups, each of the trajectory matching groups includes a first trajectory and a second trajectory, and the method further includes: determining a binary cross entropy loss based on a first difference between a first target trajectory matching group and the true trajectory of the first vehicle and the second vehicle, the first target trajectory matching group being a trajectory matching group with the highest matching degree with the true trajectory group among the trajectory matching groups, the true trajectory group including the trajectories of the first vehicle and the second vehicle; The adjusting the parameters in the edge prediction module based on the negative log-likelihood loss includes: Parameters in the edge prediction module are adjusted based on the negative log-likelihood loss and the binary cross entropy loss.

15. The method according to any one of claims 1 to 14, characterized in that The method further comprises a second training step of the trajectory interaction module: Processing the plurality of first trajectories of the first vehicle and the plurality of second trajectories of the second vehicle by the trajectory interaction module to obtain a game score of each trajectory matching group respectively; The trajectory matching group corresponding to the maximum game score is used as the second target trajectory matching group; Determining one or more of a negative log-likelihood loss, a cross entropy loss, and a regularization loss based on a difference between the second target trajectory matching group and the trajectory truth group; One or more of the negative log-likelihood loss, the cross entropy loss, and the regularization loss are used to adjust parameters in the trajectory interaction module.

16. The method according to claim 15, characterized in that The step of processing the plurality of first trajectories of the first vehicle and the plurality of second trajectories of the second vehicle by the trajectory interaction module to obtain a game score of each trajectory matching group includes: Adding the true value trajectory of the first vehicle to the plurality of first trajectories to obtain a plurality of first trajectories after data enhancement; and adding the true value trajectory of the second vehicle to the plurality of second trajectories to obtain a plurality of second trajectories after data enhancement; The plurality of first trajectories after data enhancement and the plurality of second trajectories after data enhancement are processed by the trajectory interaction module to obtain a game score of each trajectory matching group.

17. A trajectory planning device, characterized in that: include: An information acquisition unit, used to acquire vehicle proxy information and road environment information, wherein the vehicle proxy information carries vehicle driving information of the target vehicle and vehicle driving information of other vehicles, wherein the other vehicles are vehicles in the same driving area as the target vehicle; A trajectory prediction unit, used to determine a plurality of first driving trajectories of the target vehicle in a future preset time period and a plurality of second driving trajectories of the other vehicles in a future preset time period, respectively, based on the vehicle proxy information and the road environment information; The game unit is used to process the plurality of first driving trajectories and the plurality of second driving trajectories by using a trained trajectory interaction module to obtain a target predicted trajectory of the target vehicle, wherein the trajectory interaction module is constructed based on a game methodology.

18. A vehicle, characterized in that: It comprises a memory and a processor coupled to each other, wherein the processor is used to execute program instructions stored in the memory to implement the trajectory planning method described in any one of claims 1 to 16.

19. A computer-readable storage medium having program instructions stored thereon, characterized in that: When the program instructions are executed by a processor, the trajectory planning method described in any one of claims 1 to 16 is implemented.

Images