A vehicle trajectory generation method and related apparatus

Abstract

The application discloses a vehicle trajectory generation method and related device, and the method comprises the following steps: in the technical scheme of the application, firstly, determining a game feature splicing result according to first current trajectory data, second current trajectory data and at least one historical trajectory data; then, determining a candidate vehicle trajectory set of a first vehicle based on a diffusion model according to the game feature splicing result; and further, screening the candidate vehicle trajectory set to obtain a target vehicle trajectory of the first vehicle. In this way, the technical scheme of the application reflects the spatial interaction relationship between the first vehicle and the second vehicle through a game state vector, avoids the overlap between the vehicle trajectory of the first vehicle and the vehicle trajectory of the second vehicle, improves the mutual exclusivity of the generated vehicle trajectory, and makes the accuracy and reliability of the subsequently generated target vehicle trajectory higher.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of vehicle driving, in particular to a vehicle trajectory generation method and related device. BACKGROUND

[0002] With the rapid development of automatic driving technology, vehicle trajectory generation technology as a core link of automatic driving system, affects whether the automatic driving vehicle can safely drive in complex traffic environment.

[0003] In related technologies, the vehicle trajectory of the automatic driving vehicle is generally generated based on a automatic driving decision planning algorithm of imitation learning, wherein the automatic driving decision planning algorithm of imitation learning refers to imitating the driving behavior of the driver of the ego vehicle, and then generating the vehicle trajectory according to the learned driving behavior. However, since the planning algorithm only considers the driving behavior of the ego vehicle, it cannot consider the driving behavior of other vehicles, which may result in the generated vehicle trajectory overlapping with the vehicle trajectory of other vehicles, and thus the mutual exclusivity of the generated vehicle trajectory is poor. SUMMARY

[0004] Based on the above problems, the present application provides a vehicle trajectory generation method and related device, aiming to improve the mutual exclusivity of the generated vehicle trajectory.

[0005] The embodiments of the present application disclose the following technical solutions:

[0006] In a first aspect, the embodiments of the present application provide a vehicle trajectory generation method, which comprises:

[0007] obtaining first current trajectory data of a first vehicle at a current time, and obtaining second current trajectory data and historical trajectory data of at least one second vehicle within a historical preset time at the current time; wherein the first vehicle is a vehicle to be generated with a vehicle trajectory, and the second vehicle is other vehicles around the first vehicle;

[0008] determining a game feature splicing result according to the first current trajectory data, the second current trajectory data and the at least one historical trajectory data; the game feature splicing result comprises a game state vector and a splicing result after splicing of a plurality of interaction intents, the game state vector represents the interaction relationship between the first vehicle and each second vehicle, the game state vector is obtained based on the first current trajectory data and the second current trajectory data, and the interaction intent represents the probability of different driving behaviors of the first vehicle and the second vehicle, and the interaction intent is obtained based on the first current trajectory data and the at least one historical trajectory data;

[0009] determining a candidate vehicle trajectory set of the first vehicle based on a diffusion model according to the game feature splicing result; the candidate vehicle trajectory set comprises at least one candidate vehicle trajectory;

[0010] The candidate vehicle trajectory set is screened to obtain a target vehicle trajectory of the first vehicle.

[0011] With reference to the first aspect, in a possible implementation manner, the determining of the game feature splicing result according to the first current trajectory data, the second current trajectory data and the at least one historical trajectory data includes:

[0012] The first current trajectory data and the at least one second current trajectory data are processed based on a spatio-temporal graph attention mechanism network to obtain a game state vector;

[0013] According to the lane topology data, the first current trajectory data and the at least one historical trajectory data, an interaction intention of the first vehicle and the at least one second vehicle is determined to obtain a plurality of interaction intentions; wherein the lane topology data is lane information within a preset radius with the first vehicle as the center at a current time;

[0014] The game state vector and the plurality of interaction intentions are spliced to obtain the game feature splicing result.

[0015] With reference to the first aspect, in a possible implementation manner, the first current trajectory data and the at least one second current trajectory data are processed based on a spatio-temporal graph attention mechanism network to obtain a game state vector at a current time, including:

[0016] According to the first current trajectory data and the at least one historical trajectory data, feature extraction is performed based on a spatio-temporal graph attention mechanism network to determine a spatio-temporal graph; the spatio-temporal graph includes a plurality of nodes and a plurality of edges, each node includes vehicle information of a corresponding vehicle, and each edge represents a spatial interaction relationship between two nodes;

[0017] The plurality of nodes in the spatio-temporal graph are weighted aggregated to obtain graph aggregation information;

[0018] The game state vector is determined according to the graph aggregation information based on a preset neural network structure.

[0019] With reference to the first aspect, in a possible implementation manner, the plurality of nodes in the spatio-temporal graph are weighted aggregated to obtain graph aggregation information, including:

[0020] Each node in the spatio-temporal graph and a corresponding neighbor node are weighted aggregated to obtain a plurality of aggregation information;

[0021] The plurality of aggregation information are spliced to obtain graph aggregation information corresponding to the spatio-temporal graph.

[0022] With reference to the first aspect, in a possible implementation manner, the interaction intention of the first vehicle and the at least one second vehicle is determined according to the lane topology data, the first current trajectory data and the at least one historical trajectory data, and a plurality of interaction intentions are obtained, including:

[0023] The lane topology data is subjected to feature extraction to obtain lane topology features;

[0024] The key vector and the value vector are determined according to the lane topology features;

[0025] The first current trajectory data and the at least one historical trajectory data are subjected to feature extraction to obtain respective trajectory features of the first vehicle and the at least one second vehicle, and a plurality of trajectory features are obtained;

[0026] The respective query vectors of the first vehicle and the at least one second vehicle are determined according to the plurality of trajectory features;

[0027] The respective interaction intentions of the first vehicle and the at least one second vehicle are determined according to the key vector, the value vector and the query vectors, and a plurality of interaction intentions are obtained.

[0028] With reference to the first aspect, in a possible implementation manner, the interaction intention of the first vehicle and the at least one second vehicle is determined according to the key vector, the value vector and the query vectors, and a plurality of interaction intentions are obtained, including:

[0029] The respective query vectors are subjected to similarity calculation with the key vector to obtain respective attention score values of the first vehicle and the at least one second vehicle;

[0030] The plurality of attention score values are subjected to normalization processing respectively to obtain respective attention weights of the first vehicle and the at least one second vehicle;

[0031] The value vector is subjected to weighted summation processing according to the plurality of attention weights respectively to obtain respective fusion vectors of the first vehicle and the at least one second vehicle;

[0032] The respective fusion vectors are subjected to normalization processing by using a normalization function to obtain respective interaction intentions of the first vehicle and the at least one second vehicle, and a plurality of interaction intentions are obtained.

[0033] With reference to the first aspect, in a possible implementation manner, the candidate vehicle trajectory set is subjected to screening to obtain a target vehicle trajectory of the first vehicle, including:

[0034] Based on the trajectory prediction network, the reaction trajectory of each second vehicle within a preset time is determined according to the candidate vehicle trajectory set and the at least one historical trajectory data;

[0035] determine an optimization probability value of each candidate vehicle trajectory based on the at least one candidate vehicle trajectory and the at least one reaction trajectory;

[0036] take the candidate vehicle trajectory corresponding to the maximum optimization probability value in the plurality of optimization probability values as the target vehicle trajectory of the first vehicle.

[0037] In a second aspect, an embodiment of the present application provides a vehicle trajectory generation device, which comprises:

[0038] an acquisition unit, configured to acquire first current trajectory data of a first vehicle at a current time, and acquire second current trajectory data and historical trajectory data of at least one second vehicle at the current time and within a preset historical time; wherein the first vehicle is a vehicle for which a vehicle trajectory is to be generated, and the second vehicle is another vehicle around the first vehicle;

[0039] a determination unit, configured to determine a game feature splicing result according to the first current trajectory data, the second current trajectory data and the at least one historical trajectory data; the game feature splicing result comprises a game state vector and a splicing result after splicing of a plurality of interaction intents, the game state vector represents an interaction relationship between the first vehicle and each second vehicle, the game state vector is obtained based on the first current trajectory data and the second current trajectory data, and the interaction intent represents a probability of different driving behaviors of the first vehicle and the second vehicle, and the interaction intent is obtained based on the first current trajectory data and the at least one historical trajectory data;

[0040] a candidate vehicle trajectory generation unit, configured to determine a candidate vehicle trajectory set of the first vehicle based on a diffusion model and according to the game feature splicing result; the candidate vehicle trajectory set comprises at least one candidate vehicle trajectory;

[0041] a target vehicle trajectory generation unit, configured to filter the candidate vehicle trajectory set to obtain a target vehicle trajectory of the first vehicle.

[0042] In a third aspect, an embodiment of the present application provides a control device, which comprises a processor and a memory, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to complete the vehicle trajectory generation method in the first aspect.

[0043] In a fourth aspect, an embodiment of the present application provides a computer readable storage medium, which stores a computer program, and the computer program is loaded by a processor to execute the vehicle trajectory generation method in the first aspect.

[0044] In a fifth aspect, an embodiment of the present application provides a vehicle, which comprises a controller, and the controller is used to execute programs, instructions or codes in a memory to complete the vehicle trajectory generation method in the first aspect.

[0045] Beneficial effects:

[0046] In the technical solution of the present application, first, the game feature splicing result is determined according to the first current trajectory data, the second current trajectory data and the at least one historical trajectory data, then, the candidate vehicle trajectory set of the first vehicle is determined based on the diffusion model according to the game feature splicing result, and further, the target vehicle trajectory of the first vehicle is obtained by screening the candidate vehicle trajectory set.

[0047] Among them, the first vehicle is a vehicle to be generated vehicle trajectory, the second vehicle is other vehicles around the first vehicle, the game feature splicing result includes splicing result after splicing game state vector and multiple interaction intentions, the game state vector represents the interaction relationship between the first vehicle and each second vehicle, the interaction intention represents the probability of different driving behaviors of the first vehicle and the second vehicle, the game state vector is obtained based on the first current trajectory data and the second current trajectory data, the interaction intention is obtained based on the first current trajectory data and the at least one historical trajectory data, and the candidate vehicle trajectory set includes at least one candidate vehicle trajectory.

[0048] In this way, the technical solution of the present application reflects the spatial interaction relationship between the first vehicle and the second vehicle through the game state vector, and reflects the probability of different driving behaviors of the first vehicle and the at least one second vehicle through the interaction intention, so that the candidate vehicle trajectory in the subsequently generated candidate vehicle trajectory set considers the vehicle trajectory of the second vehicle, and considers the driving behavior relationship between the first vehicle and the second vehicle, that is, through the two aspects of spatial dimension and driving dimension, the overlap between the vehicle trajectory of the first vehicle and the vehicle trajectory of the second vehicle is avoided, the mutual exclusivity of the generated vehicle trajectory is improved, and the accuracy and reliability of the subsequently generated target vehicle trajectory are high. That is, the technical solution provided by the present application can improve the mutual exclusivity of the generated vehicle trajectory on the basis of ensuring high accuracy and reliability of the generated vehicle trajectory. BRIEF DESCRIPTION OF DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiment or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0050] Figure 1 A flowchart of a vehicle trajectory generation method provided by an embodiment of the present application;

[0051] Figure 2 A structure diagram of a diffusion model provided by an embodiment of the present application;

[0052] Figure 3 FIG. 1 is a structural schematic diagram of a vehicle trajectory generation device according to an embodiment of the present application.

[0053] Figure 4 FIG. 2 is a structural schematic diagram of a control device according to an embodiment of the present application. DETAILED DESCRIPTION

[0054] With the rapid development of automatic driving technology, the vehicle trajectory generation method as the core link of the automatic driving system affects whether the vehicle can safely drive and efficiently pass in a complex traffic environment. At present, how to realize the dynamic interaction processing between vehicles has become one of the main requirements for generating vehicle trajectories, such as vehicle avoidance when changing lanes, passing order at intersections, driving path when turning, and the like.

[0055] In the related art, the vehicle trajectory can be generated based on an automatic driving decision planning algorithm of imitation learning. However, the learning mechanism of the algorithm can ignore the driving behaviors of other vehicles, which can cause the generated vehicle trajectory to overlap with the vehicle trajectories of other vehicles, and further cause poor exclusivity of the generated vehicle trajectory, for example, multiple vehicles rush to the same lane when changing lanes, low passing efficiency at intersections, and the like.

[0056] In order for those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0057] Referring to Figure 1 FIG. 3 is a flowchart of a vehicle trajectory generation method according to an embodiment of the present application.

[0058] In conjunction with Figure 1 FIG. 3, the vehicle trajectory generation method provided by the present application can include:

[0059] S101: obtaining first current trajectory data of a first vehicle at a current time, and obtaining second current trajectory data and historical trajectory data of at least one second vehicle within a historical preset time at the current time.

[0060] The first vehicle is a vehicle for which a vehicle trajectory is to be generated, that is, the first vehicle is a vehicle for which a vehicle trajectory needs to be generated among a plurality of vehicles, and can also be understood as a target vehicle. The first current trajectory data is a vehicle trajectory of the first vehicle at a current time. The vehicle trajectory means a set of position information of the vehicle changing over time in the driving process, and the vehicle trajectory can be composed of a series of time-space coordinate points arranged in time sequence, and can completely record the dynamic characteristics such as the moving path, speed, direction, etc. of the vehicle.

[0061] The second vehicle is other vehicles around the first vehicle, for example, the second vehicle can be a vehicle in front of the lane where the first vehicle is located, or a vehicle in the adjacent lane of the lane where the first vehicle is located, etc. The second current trajectory data is a vehicle trajectory of the second vehicle at a current time; the historical trajectory data is a vehicle trajectory of the second vehicle within a historical preset time, for example, a vehicle trajectory of the second vehicle within the past 3 seconds.

[0062] The historical preset time means a time interval before the current time, for example, the current time is 17:35:10, and the preset time is within 3 seconds before 17:35:10, that is, 17:35:07~17:35:10. The preset time can include a plurality of time points, in one possible implementation, the time points can be divided based on a sampling frequency, for example, the sampling frequency is 10Hz, and when the preset time is 3 seconds, the 3 seconds can be divided into 30 time points (which can also be called 30 frames).

[0063] It should be noted that by dividing the preset time into a plurality of time points, the time sequence dependence (such as acceleration / deceleration trend) and spatial relationship (such as vehicle distance change) of vehicle motion can be retained.

[0064] In one possible implementation, the first current trajectory data of the ego vehicle and the second current trajectory data and the historical trajectory data of the second vehicle around the first vehicle can be synchronously collected by one or more collection devices such as a laser radar, a camera or a millimeter wave radar deployed on the first vehicle, for example, the second current trajectory data and the historical trajectory data of the second vehicle around the first vehicle at a current time and within the past 3 seconds can be synchronously collected by the laser radar, the camera and the millimeter wave radar. The trajectory data can include vehicle position (accurate to centimeter level), vehicle speed (accuracy 0.1 m / s), vehicle heading angle (accuracy 0.1 degree), wherein the sampling frequency can be 10Hz, so as to form a time-space sequence of 30 data points (i.e. 30 time points).

[0065] In a possible implementation, in order to improve the accuracy and reliability of the collection result and facilitate subsequent processing of the first current trajectory data, the second current trajectory data, and the historical trajectory data, the first current trajectory data, the second current trajectory data, and the historical trajectory data can be preprocessed.

[0066] For example, the historical trajectory data can be subjected to outlier detection and processing, and missing data can be identified in a mask manner; and the position coordinates of the second vehicle are converted into a relative coordinate system with the first vehicle as the origin.

[0067] Outlier detection and processing refers to a processing process of eliminating error data (such as sudden change of vehicle position and speed exceeding physical limit) in the historical trajectory caused by sensor noise, communication packet loss, or target occlusion, so as to improve the accuracy and reliability of the collected data. As an example, assuming that the positions of the second vehicle A at t=0s, 1s, and 2s are (10, 5), (12, 6), and (30, 8) (unit: meter), if it is detected that the position at t=2s does not match the motion trend of the previous two seconds (the speed suddenly increases from 2 m / s to 18 m / s), it is determined to be abnormal, and at this time, (30, 8) can be deleted, and the reasonable position (14, 7) at t=2s can be obtained by linear interpolation of the data at t=0s and 1s.

[0068] The mask manner for identifying missing data is used to explicitly mark the missing data position, so as to avoid errors in subsequent processing (such as coordinate conversion and trajectory prediction) caused by null values. For example, a binary array of the same dimension as the trajectory data is generated, 1 is marked for the missing position, and 0 is marked for the normal position; then, the time stamp corresponding to the missing data is recorded, so as to facilitate subsequent interpolation or model processing.

[0069] Coordinate system conversion means converting the position of the second vehicle from a global coordinate system (such as a map coordinate) to a relative coordinate system with the second vehicle as the origin, which can simplify the process of game modeling and trajectory prediction.

[0070] It should be understood that the physical reasonableness of the data can be ensured through outlier detection and processing; the missing data position can be explicitly marked through the mask manner, supporting robust processing; and the position relationship between the vehicle position of the second vehicle and the first vehicle can be converted from a global space to a local space of the first vehicle through coordinate system conversion, reducing the complexity of subsequent processing.

[0071] S102: Determine a game feature splicing result according to the first current trajectory data, the second current trajectory data, and the at least one historical trajectory data.

[0072] The game feature splicing result includes a game state vector and a splicing result after splicing of a plurality of interaction intents.

[0073] The game state vector represents the interaction relationship between the first vehicle and each second vehicle, and is obtained based on the first current trajectory data and the second current trajectory data. Specifically, the game state vector is a high-dimensional numerical vector extracted by modeling the interaction relationship of multiple vehicles (i.e., the first vehicle and at least one second vehicle). The game state vector is used to represent the game state of all participants at the current time.

[0074] For example, the individual state is the current trajectory data (such as vehicle position, vehicle speed, and vehicle heading angle) of each vehicle (the first vehicle or the second vehicle), and the interaction relationship means the relative motion trend between the individual states (i.e., between the first vehicle and the second vehicle) (for example, whether the first vehicle and the second vehicle will collide, whether the first vehicle needs to give way to the second vehicle, and the like), that is, the game state vector can represent the interaction relationship between multiple individuals (i.e., vehicles).

[0075] The interaction intention represents the probability of different driving behaviors of the first vehicle and the second vehicle respectively, and is obtained based on the first current trajectory data and at least one historical trajectory data. Specifically, the interaction intention refers to the behavior mode that the vehicle may take in the future in the game process, which is usually represented as a prediction of the probability of driving behavior.

[0076] As an example, the driving behavior is to keep straight, give way to straight, preempt straight, turn left, and turn right, the probability value of each interaction intention ranges from 0 to 1, and the sum of the probabilities of all categories is 1; then the interaction intention can be that the first vehicle keeps straight with a probability of 0.01, gives way to straight with a probability of 0.02, preempts straight with a probability of 0.06, turns left with a probability of 0.8, and turns right with a probability of 0.01.

[0077] Keep straight means that the vehicle trajectory is consistent with the current lane direction and has no turning intention; give way to straight means that the vehicle actively slows down at an intersection or merging area to give way to straight vehicles; preempt straight means that the vehicle accelerates or maintains speed at an intersection to compete with straight vehicles for road right; left turn / right turn means that the vehicle trajectory has a clear turning trend and matches the turning exclusive lane or allowed turning area in the lane topology.

[0078] It should be noted that in the embodiments of the present application, the interaction intention of the first vehicle and the interaction intention of each second vehicle need to be determined respectively, and then a plurality of interaction intentions are obtained. For example, the number of second vehicles is 10, and then 1 (first vehicle) + 10 (second vehicles) = 11 interaction intentions need to be determined.

[0079] It should be understood that in the embodiments of the present application, the game feature splicing result can represent the spatial interaction relationship between the first vehicle and the second vehicle, and can be integrated into the subsequent vehicle trajectory determination process to avoid overlap between the vehicle trajectories of the first vehicle and the second vehicle, and improve the exclusivity of the generated vehicle trajectory. At the same time, the interaction intention of the first vehicle and the second vehicle can be used to predict the intention of the ego vehicle and surrounding vehicles, avoid collision, realize cooperative planning, and improve the safety and reliability of the vehicle trajectory.

[0080] S103: determining a candidate vehicle trajectory set of the first vehicle based on the diffusion model according to the game feature splicing result.

[0081] The candidate vehicle trajectory set includes at least one candidate vehicle trajectory.

[0082] The diffusion model (Diffusion Model) is a kind of deep learning model based on probability generation. The diffusion model is to destroy data by gradually adding noise, and then learn the denoising process in reverse to generate an initial vehicle trajectory that conforms to the real data distribution. The diffusion model mainly includes two stages of forward process (adding noise) and reverse process (denoising).

[0083] In combination with Figure 2 The structure diagram of the diffusion model shown in FIG. 1, in one possible implementation, step S103 can include:

[0084] A1: inputting the game feature splicing result into each layer of the diffusion model.

[0085] It should be understood that in the embodiments of the present application, the game feature splicing result can be mapped to a space matching the input dimension of the diffusion model through a fully connected layer, and then the game feature splicing result is input as conditional information into each layer of the diffusion model.

[0086] As an example, after splicing the 256-dimensional game state vector and the multiple interaction intentions, the game feature splicing result is mapped to a space matching the input dimension of the diffusion model through a fully connected layer, and the splicing result is input as conditional information into each layer of the diffusion model.

[0087] It should be understood that the game feature splicing result not only serves as an initial input, but also dynamically adjusts the features through adaptive instance normalization (AdaIN) or cross-attention mechanism at each layer of the diffusion model (such as the encoder-decoder skip connection of U-Net) to make the generated initial vehicle trajectory more suitable for the game scenario.

[0088] A2: noise injection.

[0089] Randomly sample noise with the same dimension as the historical trajectory data from the standard Gaussian distribution, gradually increase the noise intensity in the first 50 iteration steps, for example, linearly increase the noise intensity from 0 to the maximum value (such as σ_max=1.0), simulate the process of gradually destroying the data; gradually weaken in the last 50 iteration steps, linearly decay the noise intensity from σ_max to 0, simulate the initial stage of reverse denoising.

[0090] It should be understood that the noise injection can quickly enter the high noise stage to expand the trajectory exploration range, while fine denoising in the later stage improves the smoothness of the initial vehicle trajectory.

[0091] A3: generating at least one initial vehicle trajectory.

[0092] The diffusion model generates a vehicle trajectory through 100 steps of iteration denoising, that is, in each iteration step, the diffusion model combines the current trajectory state and conditional information to predict and remove noise, and gradually generates a vehicle trajectory (i.e., an initial vehicle trajectory) closer to the true vehicle trajectory. Among them, the trajectory state can be represented by a parameterized curve (such as a cubic spline) or a time series point (such as a state point every 0.1s), which needs to be aligned with the noise dimension.

[0093] A4: using physical constraint conditions to constrain at least one initial vehicle trajectory to obtain a candidate vehicle trajectory combination including at least one candidate vehicle trajectory.

[0094] It should be understood that in order to ensure that the initial vehicle trajectory meets the vehicle dynamics, the vehicle trajectory obtained in step A3 can be verified for constraints to obtain the initial vehicle trajectory. For example, set the vehicle trajectory acceleration constraint to ±3m / s² and the curvature constraint to ≤0.1rad / m², verify the generated vehicle trajectory for constraints, and modify it by adjusting the speed and steering angle if it does not meet the constraints, until the initial vehicle trajectory is obtained.

[0095] It should be understood that the diffusion model realizes the end-to-end generation between the game state vector and the game feature splicing result after splicing the multiple interaction intentions, and the candidate vehicle trajectory set, and combines the progressive noise injection and the physical constraint initialization, which improves the diversity and game adaptability of the generated candidate vehicle trajectory on the basis of ensuring the executability of the candidate vehicle trajectory.

[0096] In the embodiments of the present application, the candidate vehicle trajectory generated by the diffusion model is smooth and meets the physical constraints (such as vehicle dynamics in autonomous driving); at the same time, by inputting the spliced game state vector and interaction intention into the diffusion model, a candidate vehicle trajectory that meets the scene requirements can be generated, that is, the candidate vehicle trajectory takes into account the spatial interaction relationship between the first vehicle and the second vehicle, thereby improving the exclusivity, accuracy and reliability of the vehicle trajectory.

[0097] S104: filtering the candidate vehicle trajectory set to obtain a target vehicle trajectory of the first vehicle.

[0098] It should be understood that in the embodiments of the present application, the target vehicle trajectory meeting the requirements can be determined by filtering the candidate vehicle trajectory set.

[0099] The vehicle trajectory generation method provided in the embodiments of the present application reflects the spatial interaction relationship between the first vehicle and the second vehicle through the game state vector, and reflects the probability of different driving behaviors of the first vehicle and the at least one second vehicle through the interaction intention, so that the candidate vehicle trajectory in the subsequently generated candidate vehicle trajectory set considers the vehicle trajectory of the second vehicle and the driving behavior relationship between the first vehicle and the second vehicle, that is, the spatial dimension and the driving dimension are considered to avoid the overlap between the vehicle trajectory of the first vehicle and the vehicle trajectory of the second vehicle, improve the exclusivity of the generated vehicle trajectory, and make the accuracy and reliability of the subsequently generated target vehicle trajectory higher. That is, the technical solution provided in the present application can improve the exclusivity of the generated vehicle trajectory on the basis of ensuring that the accuracy and reliability of the generated vehicle trajectory are relatively high.

[0100] Based on the vehicle trajectory generation method provided in the above embodiments, in a possible implementation manner, after the target vehicle trajectory is determined, the following is further included:

[0101] B1: trajectory format conversion.

[0102] The trajectory format conversion means converting the generated target vehicle trajectory (containing information such as time, position, speed, and steering angle) into a format (such as CAN bus protocol format) that can be recognized by the control system of the first vehicle.

[0103] As an example, the target vehicle trajectory can be converted into a CAN bus message, for example, 0x123 identifier corresponds to steering instruction, and 0x456 corresponds to throttle / brake instruction. As another example, if the target vehicle trajectory is generated in a global map coordinate system, the target vehicle trajectory can be converted into a vehicle local coordinate system (with the current position of the first vehicle as the origin).

[0104] It should be understood that the trajectory format conversion can realize the decoupling of the upper algorithm and the lower hardware.

[0105] B2: output frequency control.

[0106] The output frequency control refers to setting the track output cycle frequency and adopting a data check mechanism. For example, the sending period of the track instruction is fixed at 20 Hz (50 ms) through a high-precision timer (such as a hardware timer or an RTOS task scheduling), and the data check mechanism is designed to process and check the instruction sequence, the instruction correctness, and whether to resend, thereby ensuring the accuracy and reliability of the instruction.

[0107] It should be understood that the timeliness and reliability of the control instruction are ensured through the output frequency and the data check mechanism.

[0108] B3: Feedback adjustment.

[0109] The feedback adjustment refers to the control system of the first vehicle feeding back position deviation, speed deviation, and the like, and when the position deviation is greater than a preset deviation (for example, 0.5 meters), newly executing steps S101-S104 to generate a new target vehicle track for adjustment.

[0110] It should be understood that the feedback adjustment based on the position / speed deviation enables the control system to dynamically adapt to interference (such as a sudden obstacle or controller error).

[0111] Based on the vehicle track generation method provided in the above embodiments, in a possible implementation manner, step S102 can include:

[0112] C1: processing the first current track data and the at least one second current track data based on a spatio-temporal graph attention mechanism network to obtain a game state vector.

[0113] The spatio-temporal graph attention mechanism network (ST-GAT) is a deep learning model combining the graph attention mechanism (GAT) and the spatio-temporal modeling capability, and the core thereof is to capture the spatial dependency between nodes in the graph structure data and the temporal dependency in the time series through dynamic attention weights, thereby realizing joint modeling and prediction of complex spatio-temporal data.

[0114] It should be understood that in the embodiments of the present application, the spatio-temporal relationship between the first vehicle and the second vehicle at the same moment can be captured through the spatio-temporal graph attention mechanism network to obtain the game state vector at each moment.

[0115] In a possible implementation manner, step C1 can include:

[0116] D1: performing feature extraction based on the spatio-temporal graph attention mechanism network and according to the first current track data and the at least one historical track data to determine a spatio-temporal graph.

[0117] The spatiotemporal graph comprises multiple nodes and edges. Each node includes vehicle information for its corresponding vehicle; that is, every vehicle has a corresponding node. In one possible implementation, each node's information includes the vehicle's position, speed, heading angle, and type. Each edge represents the spatial interaction relationship between two nodes. In one possible implementation, the existence of an edge between the corresponding nodes can be determined using the distance between the vehicles. For example, if the distance between vehicle A and vehicle B is less than or equal to 15 meters, an edge exists between the nodes corresponding to vehicle A and vehicle B; if the distance between vehicle A and vehicle B is greater than 15 meters, no edge exists between the nodes corresponding to vehicle A and vehicle B.

[0118] It should be understood that in the embodiments of this application, the edges of the spatiotemporal graph are dynamic, and the edge weights can be updated in real time as the distance between vehicles changes, reflecting the dynamics of the scene topology.

[0119] D2: Perform weighted aggregation on multiple nodes in the spatiotemporal graph to obtain graph aggregation information.

[0120] Weighted aggregation is a core data processing method in machine learning and deep learning. It involves aggregating (merging or summarizing) multiple data sources (such as node features or time step information) by assigning different weights based on their importance or relevance, and then performing a weighted sum or combination. Its core purpose is to highlight key information, suppress noise, and improve the model's ability to capture important features.

[0121] In one possible implementation, step D2 may include: performing weighted aggregation on each node in the spatiotemporal graph and its corresponding neighbor nodes to obtain multiple aggregation information, and concatenating the multiple aggregation information to obtain the graph aggregation information corresponding to the spatiotemporal graph.

[0122] In this context, a neighboring node refers to a node that has an edge connection with the node currently being processed.

[0123] As an example, the first vehicle and at least one second vehicle are treated as nodes in a spatiotemporal graph. Vehicle information (e.g., real-time vehicle location, speed, heading angle, and vehicle type) is used as the node information for the corresponding node. An attention mechanism is used to weighted aggregate the information of each node's neighboring nodes. Assuming the first... The aggregated information of each node is represented as follows: The formula is:

[0124] ;Formula (1)

[0125] in, This represents the node information of the k-th node. Indicates the relationship with the first a set of numbers of neighbor nodes connected by edges between the nodes, and i is an integer greater than or equal to 2.

[0126] wherein, after determining the aggregated information of the plurality of nodes, the aggregated information of the plurality of nodes can be spliced to obtain the graph aggregated information .

[0127] D3: determining the game state vector based on the preset neural network structure and the graph aggregated information.

[0128] As an example, assuming that the graph aggregated information is , the graph aggregated information G can be taken as the input of the multi-layer perception machine, and then a high-dimensional game state vector is obtained through feature coding conversion, wherein a corresponding game state vector is generated at each current time.

[0129] It should be understood that, in the embodiments of the present application, since the edges of the space-time graph are dynamically generated or deleted, the rapid changes of the traffic scene can be adapted, and then the influence weight between vehicles (such as “the front vehicle” is more important than “the side and rear vehicle”) can be learned, and finally the output game state vector can reflect the complex game relationship (such as the implicit signal of “the ego vehicle wants to change lanes, and the front vehicle slows down to give way”), providing rich information for subsequent initial vehicle trajectory generation.

[0130] C2: determining the interaction intention of the first vehicle and the at least one second vehicle according to the lane topology data, the first current trajectory data and the at least one historical trajectory data, to obtain a plurality of interaction intentions.

[0131] The lane topology data means a set of lane information in a preset time range, wherein the lane topology data can be the lane information within a preset radius with the first vehicle as the center at each time. For example, the lane topology data can include lane line scatter point coordinates and adjacent connection relationship, wherein the lane line scatter point coordinates represent the lane line geometry (such as straight, curve, fork) within a preset range with the position of the first vehicle as the origin (i.e. within a preset radius with the first vehicle as the center); the adjacent connection relationship represents the topological connection between the lane lines (such as “lane A can turn left to lane B”).

[0132] In a possible implementation, step C2 can include:

[0133] E1: performing feature extraction on the lane topology data to obtain lane topology features.

[0134] It should be understood that, in the embodiments of the present application, the lane topology data can be encoded into high-dimensional lane topology features (such as 256 dimensions) in order to represent the static environmental constraints (for example, “there is a fork 50 meters ahead”) through the lane topology features.

[0135] E2: determining the key vector and the value vector according to the lane topology feature.

[0136] The key vector is used to calculate attention weight with the query vector, and the matching degree of the trajectory feature and the lane topology is measured; and the value vector is used to aggregate the relevant lane information according to the attention weight, and the environmental context is provided for the intention prediction.

[0137] E3: performing feature extraction on the first current trajectory data and the at least one historical trajectory data to obtain trajectory features corresponding to the first vehicle and the at least one second vehicle respectively, and obtaining a plurality of trajectory features.

[0138] It should be understood that in the embodiments of the present application, the first current trajectory data and the at least one historical trajectory data can be input into a multi-layer perception (MLP) to be encoded into high-dimensional trajectory features (such as 256 dimensions), so as to represent the dynamic behavior mode of the first vehicle (such as "rapid acceleration followed by deceleration" may imply a lane change intention) and the dynamic behavior mode of the second vehicle through the trajectory features.

[0139] E4: determining query vectors corresponding to the first vehicle and the at least one second vehicle according to the plurality of trajectory features.

[0140] It should be understood that in the embodiments of the present application, the query vectors corresponding to the trajectory features can be used as query conditions of the attention mechanism for subsequent screening of relevant information from the lane topology features.

[0141] E5: determining interaction intentions corresponding to the first vehicle and the at least one second vehicle according to the key vector, the value vector and the query vector, and obtaining a plurality of interaction intentions.

[0142] It should be understood that in the embodiments of the present application, the motion trend of the vehicle is reflected through the trajectory features, and the scene constraint is provided through the lane topology features (such as "vehicle in straight lane but deceleration" may be yielding rather than left turn), which can make the subsequent determined interaction intention more accurate.

[0143] In one possible implementation, step E5 can include:

[0144] F1: performing similarity calculation on each query vector and the key vector respectively to obtain attention score values corresponding to the first vehicle and the at least one second vehicle respectively.

[0145] The attention score value represents the matching degree of the trajectory feature and the lane topology feature (such as the dot product value of "trajectory left offset" and "left turn lane line" is high), so as to dynamically screen the lane topology feature most relevant to the trajectory feature.

[0146] F2: normalize the plurality of attention score values respectively to obtain the attention weight corresponding to the first vehicle and the at least one second vehicle respectively.

[0147] It should be understood that in the embodiments of the present application, by normalizing to determine the attention weight corresponding to the first vehicle and the at least one second vehicle respectively, the low correlation lane topology features (weight close to 0) can be inhibited, and the high correlation lane topology features (weight close to 1) can be highlighted.

[0148] It should be understood that by steps F1 and F2, the weight between the trajectory feature and the lane topology feature can be dynamically allocated (such as "more dependent on trajectory when there is no lane line"), which is beneficial to improve the accuracy and reliability of the interaction intention prediction.

[0149] F3: according to the plurality of attention weights, weighted sum processing is performed on the value vectors respectively to obtain the fusion vector corresponding to the first vehicle and the at least one second vehicle respectively.

[0150] It should be understood that by performing weighted sum processing on the value vectors according to the attention weights, the information in the lane topology feature that is most relevant to the trajectory feature can be combined (such as "left turn lane line position + adjacent lane yield rule"), and the interference of irrelevant branches (such as right turn lane) can be avoided.

[0151] F4: normalizing each fusion vector by using a normalization function to obtain the interaction intention corresponding to the first vehicle and the at least one second vehicle respectively, and obtaining a plurality of interaction intentions.

[0152] The normalization function means a function used for predicting the interaction intention corresponding to the first vehicle and the at least one second vehicle respectively. In a possible implementation, the normalization function can be a Softmax activation function. The Softmax activation function (Softmax Function) is a mathematical tool for mapping a plurality of values to a probability distribution, and is commonly used in the output layer of a multi-classification task or an attention mechanism. In the embodiments of the present application, the normalization function can convert the fusion vector into a probability distribution, so as to represent the probability of the driving behavior of the vehicle at the corresponding moment.

[0153] C3: splicing the game state vector and the plurality of interaction intentions to obtain a game feature splicing result.

[0154] It should be understood that in the embodiments of the present application, by splicing the game state vector and the plurality of interaction intentions to obtain the game feature splicing result, the game feature splicing result can be used as conditional information to provide scene context for the diffusion model, so that the trajectory generated by the diffusion model conforms to the current game state and intention.

[0155] Based on the vehicle trajectory generation method provided in the above embodiments, in order to determine the target vehicle trajectory, the game strategy screening of safe and comfortable guidance needs to be carried out, and therefore the trajectory can be optimized from the aspects of game iteration optimization and accelerated sampling and constraint verification in the embodiments of the present application. In a possible implementation manner, step S104 can include the following steps.

[0156] G1: Based on the trajectory prediction network, the reaction trajectory of each second vehicle within a preset time is determined according to the candidate vehicle trajectory set and at least one historical trajectory data.

[0157] It should be understood that the generated candidate trajectory of the ego vehicle is used as one of the input conditions, and the trajectory prediction generation network (inputting the past 3-second trajectory data of the surrounding vehicle and the candidate trajectory of the ego vehicle, containing a 128-dimensional hidden layer) is used to predict the reaction trajectory of the surrounding vehicle in the future 5 seconds.

[0158] G2: The optimization probability value of each candidate vehicle trajectory is determined based on at least one candidate vehicle trajectory and at least one reaction trajectory.

[0159] As an example, the optimization probability value can be obtained in the following manner:

[0160] The safety constraint term of each candidate vehicle trajectory is calculated :

[0161] ; formula (2)

[0162] Wherein, the default value of σ is 0.5, represents the minimum distance between the ith candidate trajectory and the reaction trajectory of the jth surrounding vehicle.

[0163] Then, the comfort constraint term of each candidate trajectory corresponding to the vehicle state data of the first vehicle is determined. The vehicle state data means the state parameters of the first vehicle in the driving process, such as brake parameters, speed parameters, etc., which are not limited here.

[0164] Further, the optimization probability value is obtained according to the weighted sum of the safety constraint term and the comfort constraint term .

[0165] G3: The candidate vehicle trajectory corresponding to the maximum optimization probability value in the plurality of optimization probability values is taken as the target vehicle trajectory of the first vehicle.

[0166] It should be understood that the candidate vehicle trajectory with the maximum optimization probability value is taken as the target vehicle trajectory, which can ensure that the target vehicle trajectory meets the requirements of safety and comfort.

[0167] Based on the vehicle trajectory generation method provided in the above embodiments, referring to Figure 3 FIG. 2 is a structural schematic diagram of a vehicle trajectory generation device provided in an embodiment of the present application.

[0168] In combination with Figure 3 As shown in the figure, the vehicle trajectory generation device 30 provided in the embodiment of the present application can include:

[0169] The acquisition unit 31 is configured to acquire first current trajectory data of a first vehicle at a current time, and acquire second current trajectory data and historical trajectory data of at least one second vehicle at the current time and within a preset historical time; wherein the first vehicle is a vehicle for which a vehicle trajectory is to be generated, and the second vehicle is another vehicle around the first vehicle;

[0170] The determination unit 32 is configured to determine a game feature splicing result according to the first current trajectory data, the second current trajectory data and the at least one historical trajectory data; the game feature splicing result includes a splicing result after splicing of a game state vector and a plurality of interaction intents, the game state vector represents an interaction relationship between the first vehicle and each second vehicle, the game state vector is obtained based on the first current trajectory data and the second current trajectory data, and the interaction intent represents a probability of different driving behaviors of the first vehicle and the second vehicle, and the interaction intent is obtained based on the first current trajectory data and the at least one historical trajectory data;

[0171] The candidate vehicle trajectory generation unit 33 is configured to determine a candidate vehicle trajectory set of the first vehicle according to the game feature splicing result based on a diffusion model; the candidate vehicle trajectory set includes at least one candidate vehicle trajectory;

[0172] The target vehicle trajectory generation unit 34 is configured to filter the candidate vehicle trajectory set to obtain a target vehicle trajectory of the first vehicle.

[0173] In a possible implementation, the determination unit 32 includes:

[0174] The first determination subunit is configured to process the first current trajectory data and the at least one second current trajectory data based on a spatio-temporal graph attention mechanism network to obtain the game state vector;

[0175] The second determination subunit is configured to determine the interaction intents of the first vehicle and the at least one second vehicle according to lane topology data, the first current trajectory data and the at least one historical trajectory data to obtain the plurality of interaction intents; wherein the lane topology data is lane information within a preset radius with the first vehicle as the center at the current time;

[0176] The third determination subunit is configured to splice the game state vector and the plurality of interaction intents to obtain the game feature splicing result.

[0177] In a possible implementation, the first determining subunit is configured to:

[0178] perform feature extraction on the first current trajectory data and the at least one historical trajectory data based on a spatio-temporal graph attention mechanism network to determine a spatio-temporal graph; the spatio-temporal graph includes a plurality of nodes and a plurality of edges, each node includes vehicle information of a corresponding vehicle, and each edge represents a spatial interaction relationship between two nodes;

[0179] perform weighted aggregation on the plurality of nodes in the spatio-temporal graph to obtain graph aggregation information;

[0180] determine a game state vector based on the graph aggregation information based on a preset neural network structure.

[0181] In a possible implementation, the first determining subunit is specifically configured to:

[0182] perform weighted aggregation between each node in the spatio-temporal graph and a corresponding neighbor node to obtain a plurality of aggregation information;

[0183] concatenate the plurality of aggregation information to obtain graph aggregation information corresponding to the spatio-temporal graph.

[0184] In a possible implementation, the second determining subunit is configured to:

[0185] perform feature extraction on the lane topology data to obtain lane topology features;

[0186] determine a key vector and a value vector based on the lane topology features;

[0187] perform feature extraction on the first current trajectory data and the at least one historical trajectory data to obtain trajectory features corresponding to the first vehicle and the at least one second vehicle respectively, and obtain a plurality of trajectory features;

[0188] determine a query vector corresponding to the first vehicle and the at least one second vehicle respectively based on the plurality of trajectory features;

[0189] determine an interaction intention corresponding to the first vehicle and the at least one second vehicle respectively based on the key vector, the value vector, and the query vector, and obtain a plurality of interaction intentions.

[0190] In a possible implementation, the second determining subunit is specifically configured to:

[0191] perform similarity calculation on each query vector and the key vector respectively to obtain an attention score value corresponding to the first vehicle and the at least one second vehicle respectively;

[0192] perform normalization processing on the plurality of attention score values respectively to obtain an attention weight corresponding to the first vehicle and the at least one second vehicle respectively;

[0193] The value vectors are respectively weighted and summed according to the plurality of attention weights to obtain a fusion vector corresponding to each of the first vehicle and the at least one second vehicle;

[0194] The normalization function is used to normalize each fusion vector to obtain an interaction intention corresponding to each of the first vehicle and the at least one second vehicle, and a plurality of interaction intentions are obtained.

[0195] In a possible implementation, the target vehicle trajectory generation unit 34 is configured to:

[0196] Based on the trajectory prediction network, the reaction trajectory of each second vehicle within a preset time is determined according to the candidate vehicle trajectory set and the at least one historical trajectory data;

[0197] Based on the at least one candidate vehicle trajectory and the at least one reaction trajectory, an optimization probability value of each candidate vehicle trajectory is determined;

[0198] The candidate vehicle trajectory corresponding to the maximum optimization probability value in the plurality of optimization probability values is taken as the target vehicle trajectory of the first vehicle.

[0199] The vehicle trajectory generation device provided by the embodiments of the present application has the same beneficial effects as the vehicle trajectory generation method provided by the above embodiments, and thus will not be described again.

[0200] In a possible implementation, referring to Figure 4 The figure is a schematic diagram of a control device provided by an embodiment of the present application.

[0201] The control device can include a memory 411 and a processor 412. As Figure 4 shown, the memory can be a random access memory (RAM), a flash memory, a read-only memory (ROM), an EPROM memory, a non-volatile read-only memory (Electronic Programmable ROM, EPROM), a register, a hard disk, a removable disk, etc.

[0202] The memory 411 can store computer instructions, and when the computer instructions stored in the memory 411 are executed by the processor 412, the processor 412 can be configured to execute the vehicle trajectory generation method. The memory 411 can also store data, such as the preset time, the preset time, the preset constraint condition and other information involved in the above embodiments.

[0203] In the above embodiments, all or part of the embodiments can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions according to the embodiments of the present application are generated. The computer can be a general purpose computer, a special purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer readable storage medium or transmitted from one computer readable storage medium to another computer readable storage medium, for example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center through wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) mode. The computer readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. integrated with one or more available media. The available media can be magnetic media (such as floppy disk, hard disk, magnetic tape), or semiconductor media (such as solid state disk (SSD)) and the like.

[0204] The embodiments of the present application also provide a readable storage medium for storing the method provided by the above embodiments. For example, random access memory (RAM), flash memory, read-only memory (ROM), EPROM memory, non-volatile read-only memory (EPROM), register, hard disk, removable disk or any other form of storage medium in the art.

[0205] The "first", "second" in the names mentioned in the embodiments of the present application are only used for name identification, and do not represent the first and second in order.

[0206] It should be noted that the embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts of each embodiment can be referred to each other. For the method disclosed in the embodiments, since it corresponds to the product embodiment disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the product embodiment.

[0207] The foregoing description of the disclosed embodiments enables a person skilled in the art to make or use the application. Modifications of these embodiments will occur to persons of skill in the art, and that the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Therefore, the present application is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A vehicle trajectory generation method characterized by, The method comprises: obtaining first current trajectory data of a first vehicle at a current time, and obtaining second current trajectory data and historical trajectory data of at least one second vehicle at the current time and within a preset historical time; wherein the first vehicle is a vehicle to be generated with a vehicle trajectory, and the second vehicle is other vehicles around the first vehicle; determining a game feature splicing result according to the first current trajectory data, the second current trajectory data and at least one historical trajectory data; the game feature splicing result comprises a game state vector and a splicing result after splicing of multiple interaction intents, the game state vector represents an interaction relationship between the first vehicle and each second vehicle, the game state vector is obtained based on the first current trajectory data and the second current trajectory data, and the interaction intent represents a probability of different driving behaviors of the first vehicle and at least one second vehicle, and the interaction intent is obtained based on the first current trajectory data and at least one historical trajectory data; determining a candidate vehicle trajectory set of the first vehicle based on a diffusion model according to the game feature splicing result; the candidate vehicle trajectory set comprises at least one candidate vehicle trajectory; screening the candidate vehicle trajectory set to obtain a target vehicle trajectory of the first vehicle.

2. The vehicle trajectory generation method according to claim 1, characterized by, The game feature splicing result is determined according to the first current trajectory data, the second current trajectory data and at least one historical trajectory data, comprising: processing the first current trajectory data and at least one second current trajectory data based on a spatio-temporal graph attention mechanism network to obtain a game state vector; determining interaction intents of the first vehicle and at least one second vehicle according to lane topology data, the first current trajectory data and at least one historical trajectory data to obtain multiple interaction intents; wherein the lane topology data is lane information within a preset radius with the first vehicle as the center at the current time; splicing the game state vector and multiple interaction intents to obtain the game feature splicing result.

3. The vehicle trajectory generation method according to claim 2, characterized by, The game state vector is obtained by processing the first current trajectory data and at least one second current trajectory data based on a spatio-temporal graph attention mechanism network, comprising: extracting features based on a spatio-temporal graph attention mechanism network according to the first current trajectory data and at least one historical trajectory data to determine a spatio-temporal graph; the spatio-temporal graph comprises multiple nodes and multiple edges, each node comprises vehicle information of a corresponding vehicle, and each edge represents a spatial interaction relationship between two nodes; weighting and aggregating multiple nodes in the spatio-temporal graph to obtain graph aggregation information; determining the game state vector based on a preset neural network structure according to the graph aggregation information.

4. The vehicle trajectory generation method according to claim 3, characterized by, The multiple nodes in the spatio-temporal graph are weighted and aggregated to obtain the graph aggregation information, comprising: weighting and aggregating each node in the spatio-temporal graph and the corresponding neighbor node to obtain multiple aggregation information; splicing multiple aggregation information to obtain graph aggregation information corresponding to the spatio-temporal graph.

5. The vehicle trajectory generation method according to claim 2, characterized by, determine interaction intentions of the first vehicle and the at least one second vehicle according to the lane topology data, the first current trajectory data, and the at least one historical trajectory data, to obtain a plurality of interaction intentions, including: perform feature extraction on the lane topology data to obtain lane topology features; determine a key vector and a value vector according to the lane topology features; perform feature extraction on the first current trajectory data and the at least one historical trajectory data to obtain trajectory features corresponding to the first vehicle and the at least one second vehicle respectively, to obtain a plurality of trajectory features; determine query vectors corresponding to the first vehicle and the at least one second vehicle respectively according to the plurality of trajectory features; determine interaction intentions corresponding to the first vehicle and the at least one second vehicle respectively according to the key vector, the value vector, and the query vectors, to obtain the plurality of interaction intentions.

6. The vehicle trajectory generation method according to claim 5, characterized by, The determining of the interaction intentions corresponding to the first vehicle and the at least one second vehicle respectively according to the key vector, the value vector, and the query vectors to obtain the plurality of interaction intentions includes: perform similarity calculation on each of the query vectors and the key vector respectively to obtain attention score values corresponding to the first vehicle and the at least one second vehicle respectively; perform normalization processing on a plurality of the attention score values respectively to obtain attention weights corresponding to the first vehicle and the at least one second vehicle respectively; perform weighted summation processing on the value vector according to a plurality of the attention weights respectively to obtain fusion vectors corresponding to the first vehicle and the at least one second vehicle respectively; perform normalization processing on each of the fusion vectors by using a normalization function to obtain the interaction intentions corresponding to the first vehicle and the at least one second vehicle respectively, to obtain the plurality of interaction intentions.

7. The vehicle trajectory generation method according to claim 1, characterized by, The filtering of the candidate vehicle trajectory set to obtain the target vehicle trajectory of the first vehicle includes: determine reaction trajectories of each of the second vehicles within a preset time according to the candidate vehicle trajectory set and the at least one historical trajectory data based on a trajectory prediction network; determine optimization probability values of each of the candidate vehicle trajectories based on the at least one candidate vehicle trajectory and the at least one reaction trajectory; take a candidate vehicle trajectory corresponding to a maximum optimization probability value in a plurality of the optimization probability values as the target vehicle trajectory of the first vehicle.

8. A vehicle trajectory generating apparatus characterized by comprising: The apparatus includes: an acquisition unit configured to acquire first current trajectory data of a first vehicle at a current time, and to acquire second current trajectory data and historical trajectory data within a historical preset time of at least one second vehicle at the current time; wherein the first vehicle is a vehicle to be generated with a vehicle trajectory, and the second vehicle is another vehicle around the first vehicle; determining unit configured to determine a game feature splicing result according to the first current trajectory data, the second current trajectory data, and the at least one historical trajectory data; the game feature splicing result includes a splicing result of a game state vector and a plurality of interaction intents, the game state vector represents an interaction relationship between the first vehicle and each of the second vehicles, the game state vector is obtained based on the first current trajectory data and the second current trajectory data, and the interaction intent represents a probability of different driving behaviors of the first vehicle and the second vehicles, the interaction intent is obtained based on the first current trajectory data and the at least one historical trajectory data; a candidate vehicle trajectory generation unit configured to determine a candidate vehicle trajectory set of the first vehicle based on a diffusion model and according to the game feature splicing result; the candidate vehicle trajectory set includes at least one candidate vehicle trajectory; a target vehicle trajectory generation unit configured to screen the candidate vehicle trajectory set to obtain a target vehicle trajectory of the first vehicle.

9. A control device characterized by comprising: A vehicle trajectory generation method includes a processor and a memory, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to complete the vehicle trajectory generation method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, A computer program is stored, and the computer program is loaded by a processor to execute the vehicle trajectory generation method according to any one of claims 1-7.

Images