Trajectory prediction method, device and equipment for autonomous vehicle, and medium

Abstract

The invention is suitable for the technical field of automatic driving, and provides a trajectory prediction method and device for an automatic driving vehicle, equipment and a medium, and the method comprises the steps: obtaining scene map data of a target scene where a target vehicle is located, historical trajectory data of a plurality of intelligent vehicles in the target scene, and a current driving state of the target vehicle; according to the historical trajectory data and the scene map data of the plurality of intelligent vehicles, determining a plurality of candidate prediction trajectories of the target vehicle at a future moment and corresponding trajectory prediction probability values; calculating a risk cost value of the target vehicle according to the trajectory prediction probability values corresponding to the plurality of candidate prediction trajectories of the target vehicle at the future moment and the current driving states of the plurality of intelligent vehicles; and according to the risk cost value of the target vehicle, optimizing the candidate prediction trajectory of which the trajectory prediction probability value is the highest value, and generating a target trajectory of the target vehicle at a future moment, so as to improve the safety and adaptability of trajectory prediction of the vehicle in a complex traffic scene.

Description

Technical Field

[0001] This application belongs to the technical field of autonomous driving, and particularly relates to a method, device, equipment and medium for predicting the trajectory of an autonomous vehicle. Background Art

[0002] At present, the trajectory prediction of vehicles is crucial in an autonomous driving system and is a core part of the autonomous driving system architecture. An autonomous vehicle can accurately predict the future trajectories of surrounding vehicles, pedestrians and other traffic participants, and predict and plan an obstacle avoidance path in advance.

[0003] However, in a complex traffic environment, the existing trajectory prediction of autonomous vehicles only makes a passive response to the prediction of other vehicles, and there are problems of poor safety and adaptability in trajectory prediction. Summary of the Invention

[0004] The embodiments of this application provide a method, device, equipment and medium for predicting the trajectory of an autonomous vehicle, which can solve the problems of poor safety and adaptability in the existing trajectory prediction method of autonomous vehicles in a complex traffic environment.

[0005] In a first aspect, the embodiments of this application provide a method for predicting the trajectory of an autonomous vehicle, and the method includes:

[0006] Obtain the scene map data of the target scene where the target vehicle is located, the historical trajectory data of multiple intelligent vehicles in the target scene, and the current driving states of multiple intelligent vehicles, where the intelligent vehicles include the target vehicle and multiple surrounding vehicles;

[0007] According to the historical trajectory data of multiple intelligent vehicles and the scene map data, determine multiple candidate prediction trajectories of the target vehicle at a future moment and the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories;

[0008] According to the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories of the target vehicle at a future moment and the current driving states of multiple intelligent vehicles, calculate the risk cost value of the target vehicle;

[0009] According to the risk cost value of the target vehicle, optimize the candidate prediction trajectory with the highest trajectory prediction probability value to generate the target trajectory of the target vehicle at a future moment.

[0010] In a possible implementation manner of the first aspect, the determining, according to the historical trajectory data of multiple intelligent vehicles and the scene map data, multiple candidate prediction trajectories of the target vehicle at a future moment and the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories includes:

[0011] Encode the historical trajectory data of the multiple intelligent vehicles through a trajectory encoder to obtain the encoded historical trajectory data of the multiple intelligent vehicles;

[0012] Obtain the relative position data between the multiple intelligent vehicles according to the encoded historical trajectory data of the multiple intelligent vehicles;

[0013] Determine the interaction features of the multiple intelligent vehicles according to the scenario map data, the encoded historical trajectory data of the multiple intelligent vehicles, and the relative position data between the multiple intelligent vehicles;

[0014] Extract features from the encoded historical trajectory data through a multi-layer perceptron to determine the intention features of the multiple intelligent vehicles and the intention prediction probability values corresponding to the intention features, where the intention features include lateral intention features and longitudinal intention features;

[0015] Obtain multiple candidate prediction trajectories of the target vehicle at a future moment and trajectory prediction probability values corresponding to the multiple candidate prediction trajectories according to the interaction features of the multiple intelligent vehicles and the intention features of the multiple intelligent vehicles.

[0016] In a possible implementation manner of the first aspect, the obtaining multiple candidate prediction trajectories of the target vehicle at a future moment and trajectory prediction probability values corresponding to the multiple candidate prediction trajectories according to the interaction features of the multiple intelligent vehicles and the intention features of the multiple intelligent vehicles includes:

[0017] Fuse the interaction features of the multiple intelligent vehicles and the intention features of the multiple intelligent vehicles to obtain a fused feature;

[0018] Concatenate the fused feature and the intention prediction probability value corresponding to the intention feature, and input them into a multi-layer perceptron MLP for processing to obtain an embedding vector corresponding to the fused feature;

[0019] Input the embedding vector into a long short-term memory LSTM decoder to obtain multiple candidate prediction trajectories of the target vehicle at a future moment and trajectory prediction probability values corresponding to the multiple candidate prediction trajectories.

[0020] In a possible implementation manner of the first aspect, the determining the interaction features of the multiple intelligent vehicles according to the scenario map data, the encoded historical trajectory data of the multiple intelligent vehicles, and the relative position data between the multiple intelligent vehicles includes:

[0021] Encode the scene map data through a map encoder to obtain the encoded scene map data;

[0022] Based on the encoded historical trajectory data of multiple intelligent vehicles, obtain the first interaction features among the multiple intelligent vehicles;

[0023] Based on the encoded historical trajectory data of multiple intelligent vehicles and the encoded scene map data, obtain the second interaction features between the multiple intelligent vehicles and the map;

[0024] Based on the first interaction features, the second interaction features, and the relative position data among the multiple intelligent vehicles, obtain the interaction features of the multiple intelligent vehicles.

[0025] In a possible implementation manner of the first aspect, the calculating the risk cost value of the target vehicle according to the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories of the target vehicle at a future moment and the current driving states of the multiple intelligent vehicles includes:

[0026] Based on the current driving states of the multiple intelligent vehicles, determine the collision probability values and estimated injury values of the multiple intelligent vehicles on the candidate prediction trajectory with the highest trajectory prediction probability value;

[0027] Based on the collision probability values and the estimated injury values, obtain the risk values of each intelligent vehicle for a collision;

[0028] Based on the risk values of each intelligent vehicle for a collision, calculate the risk cost value of the target vehicle.

[0029] In a possible implementation manner of the first aspect, the calculating the risk cost value of the target vehicle according to the risk values of each intelligent vehicle for a collision includes at least one of the following:

[0030] Based on the risk values of each intelligent vehicle and a preset boundary damage value on the candidate prediction trajectory with the highest trajectory prediction probability value, calculate the self - protection risk cost value of the target vehicle;

[0031] Based on the difference between the risk values of any two intelligent vehicles on the candidate prediction trajectory with the highest trajectory prediction probability value, calculate the caring cost value of the target vehicle;

[0032] Based on the risk values of each intelligent vehicle and a preset scaling factor on the candidate prediction trajectory with the highest trajectory prediction probability value, calculate the sudden high - risk cost value of the target vehicle;

[0033] Calculate the risk cost value of the target vehicle based on the self - protection risk cost value, the care cost value, and the sudden high - risk cost value of the target vehicle.

[0034] In a possible implementation manner of the first aspect, the method further includes:

[0035] Based on a preset loss function, calculate the loss value between the target trajectory and the candidate prediction trajectories, where the loss value includes the smooth loss value of displacement deviation and the cross - entropy loss value of intention recognition;

[0036] Optimize the risk optimization module according to the risk cost value, the smooth loss value of displacement deviation, and the cross - entropy loss value of intention recognition, to obtain the optimized risk optimization module, where the risk optimization module is used to optimize the candidate prediction trajectory with the highest trajectory prediction probability value.

[0037] In a second aspect, an embodiment of the present application provides a trajectory prediction device for an autonomous driving vehicle, the device includes:

[0038] An acquisition module, configured to acquire the scene map data of the target scene where the target vehicle is located, the historical trajectory data of multiple intelligent vehicles in the target scene, and the current driving state of the target vehicle, where the intelligent vehicles include the target vehicle and multiple surrounding vehicles;

[0039] A prediction module, configured to determine multiple candidate prediction trajectories of the target vehicle at a future moment and the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories according to the historical trajectory data of the multiple intelligent vehicles and the scene map data; [[ID=२०]]

[0040] A calculation module, configured to calculate the risk cost value of the target vehicle according to the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories of the target vehicle at a future moment and the current driving states of the multiple intelligent vehicles;

[0041] An optimization module, configured to optimize the candidate prediction trajectory with the highest trajectory prediction probability value according to the risk cost value of the target vehicle, and generate the target prediction trajectory of the target vehicle at a future moment.

[0042] In a third aspect, an embodiment of the present application provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, where when the processor executes the computer program, it implements the trajectory prediction method of the autonomous driving vehicle described in any one of the above.

[0043] Fourthly, an embodiment of the present application provides a computer-readable storage medium storing a computer program, which when executed by a processor implements the trajectory prediction method for an autonomous vehicle described in any one of the above.

[0044] Fifthly, an embodiment of the present application provides a computer program product, which when running on a terminal device causes the terminal device to execute the trajectory prediction method for an autonomous vehicle described in any one of the first aspects above.

[0045] The beneficial effects of the embodiments of the present application compared with the prior art are as follows:

[0046] An embodiment of the present application provides a trajectory prediction method for an autonomous vehicle. The method includes: obtaining scene map data of a target scene where a target vehicle is located, historical trajectory data of multiple intelligent vehicles in the target scene, and the current driving state of the target vehicle, where the intelligent vehicles include the target vehicle and multiple surrounding vehicles; determining multiple candidate prediction trajectories of the target vehicle at a future moment and trajectory prediction probability values corresponding to the multiple candidate prediction trajectories according to the historical trajectory data of the multiple intelligent vehicles and the scene map data; calculating a risk cost value of the target vehicle according to the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories of the target vehicle at the future moment and the current driving states of the multiple intelligent vehicles; and optimizing the candidate prediction trajectory with the highest trajectory prediction probability value according to the risk cost value of the target vehicle to generate a target trajectory of the target vehicle at the future moment. The present application generates multiple candidate prediction trajectories of the target vehicle through the historical trajectory data of the intelligent vehicles and the target scene map data, and optimizes the candidate prediction trajectories through the calculated risk cost value of the target vehicle to improve the safety and adaptability of the trajectory prediction of the intelligent vehicle in a complex traffic scene. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] To more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for use in the embodiments or the description of the prior art. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0048] Figure 1 FIG. is a schematic diagram of the process of trajectory generation to optimization of an autonomous vehicle in a variety of vehicle scenarios provided by an embodiment of the present application;

[0049] Figure 2 FIG. is a schematic flowchart of a trajectory prediction method for an autonomous vehicle provided by an embodiment of the present application;

[0050] Figure 3It is a schematic flowchart of a trajectory prediction method for an autonomous vehicle provided by another embodiment of the present application;

[0051] Figure 4 It is a schematic structural diagram of a trajectory prediction device for an autonomous vehicle provided by an embodiment of the present application;

[0052] Figure 5 It is a schematic structural diagram of a terminal device provided by an embodiment of the present application. Detailed implementation manners

[0053] In the following description, for the purpose of illustration rather than limitation, specific details such as specific system architectures and technologies are presented in order to thoroughly understand the embodiments of the present application. However, those skilled in the art should clearly understand that the present application can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid unnecessary details from interfering with the description of the present application.

[0054] It should be understood that when used in the specification and appended claims of the present application, the term "comprising" indicates the presence of the described features, wholes, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components, and / or their combinations.

[0055] It should also be understood that the term "and / or" as used in the specification and appended claims of the present application refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0056] As used in the specification and appended claims of the present application, the term "if" can be interpreted as "when", "once", "in response to determining", or "in response to detecting" depending on the context. Similarly, the phrase "if determined" or "if detected [the described condition or event]" can be interpreted as meaning "once determined", "in response to determining", "once detected [the described condition or event]", or "in response to detecting [the described condition or event]" depending on the context.

[0057] In addition, in the description of the specification and appended claims of the present application, the terms "first", "second", "third", etc. are only used for distinguishing descriptions and cannot be understood as indicating or implying relative importance.

[0058] References to "an embodiment" or "some embodiments" etc. described in the specification of this application mean that a particular feature, structure, or characteristic described in connection with the embodiment is included in one or more embodiments of this application. Thus, statements such as "in one embodiment", "in some embodiments", "in other some embodiments", "in still other embodiments", etc. that appear in different places in this specification do not necessarily all refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in another way. The terms "comprising", "including", "having" and their variants all mean "including but not limited to", unless otherwise specifically emphasized in another way.

[0059] In a complex traffic environment, it remains a major challenge for autonomous vehicles to accurately predict the trajectories of surrounding vehicles as well as human drivers. The behavior of a vehicle is affected not only by its historical motion but also by the behavior of surrounding vehicles. To address this issue, the prior art has proposed a Conditional Marginal Prediction (CMP) model that predicts the future trajectories of other vehicles based on the future trajectories of the queried vehicle. However, the drawback of the CMP model is that even in critical situations such as merging, lane changing, or unprotected left turns, autonomous vehicles can only passively react to the predicted behavior of other vehicles. However, in these situations, autonomous vehicles and other vehicles should actively coordinate rather than just passively react to the prediction.

[0060] In addition, due to the inherent randomness and uncertainty in driver behavior, there may be multiple reasonable trajectory choices even in the same situation. Faced with this uncertainty, traditional prediction models usually generate multiple possible predicted trajectories, which increases the complexity of decision-making. However, intent recognition can effectively reduce this uncertainty by identifying the most likely driving behavior or trajectory in a given situation. Therefore, the model proposed herein prioritizes intent-based trajectory prediction rather than considering all possible motion patterns, making decision-making more efficient and accurate.

[0061] Finally, in complex mixed traffic scenarios, especially those with collision risks, the decision-making strategy must focus on accident avoidance. In the prior art, it can be roughly divided into classical methods and learning-based methods. Among them, classical methods, such as Model Predictive Control (MPC) combined with potential field technology, provide reasonable decisions for autonomous vehicles, but there is a problem of relying on fixed parameters; while learning-based methods can learn appropriate decisions, but the scene generalization ability is poor. In contrast, the trajectory prediction method for autonomous vehicles provided in this application selects the concept of risk allocation to evaluate the generated trajectories, which not only realizes a reasonable decision-making strategy but also enhances the robustness of the model.

[0062] Therefore, this application models the interaction between vehicles based on the concept of joint prediction, solves the uncertainty problem by introducing intention recognition based on driving behavior, and then constrains the generated trajectories. As shown in the figure, Figure 1 is a schematic diagram of the process from trajectory generation to optimization of an autonomous vehicle in multiple vehicle scenarios provided by an embodiment of this application. Figure 1 In it, serial number Ⅰ represents the autonomous vehicle, and serial number Ⅱ represents other vehicles, illustrating the process of the trajectory of the autonomous vehicle from generation to optimization. In addition, by applying the risk allocation principle, the adaptability in complex traffic scenarios is improved. If it is detected that the trajectory of the autonomous vehicle intersects with high-risk areas such as sidewalks, the predicted trajectory is optimized to improve the safety and adaptability of the decision-making and planned trajectory of the autonomous vehicle in complex scenarios.

[0063] Please refer to Figure 2 , Figure 2 is a schematic flowchart of a trajectory prediction method for an autonomous vehicle provided by an embodiment of this application. By way of example and not limitation, this method can be applied to terminal devices such as servers, such as autonomous vehicles. This method includes:

[0064] S11. Obtain the scene map data of the target scene where the target vehicle is located, the historical trajectory data of multiple intelligent vehicles in the target scene, and the current driving states of the multiple intelligent vehicles.

[0065] Among them, the intelligent vehicles include the target vehicle and multiple surrounding vehicles.

[0066] S12. Determine multiple candidate prediction trajectories of the target vehicle at future moments and the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories according to the historical trajectory data of the multiple intelligent vehicles and the scene map data.

[0067] S13. Calculate the risk cost value of the target vehicle according to the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories of the target vehicle at future moments and the current driving states of the multiple intelligent vehicles.

[0068] S14. Optimize the candidate prediction trajectory with the highest trajectory prediction probability value according to the risk cost value of the target vehicle, and generate the target trajectory of the target vehicle at a future moment.

[0069] An intelligent vehicle is a vehicle with autonomous driving or assisted driving functions traveling in a target scenario, which may include the target vehicle and surrounding vehicles related to the target vehicle. The target vehicle may be an autonomous driving vehicle that needs trajectory prediction. In this embodiment, the target vehicle may be used as the host vehicle.

[0070] The target scenario may be the specific environment or scenario where the target vehicle is located, and the target scenario may include scenarios such as roads, intersections, and parking lots. The scenario map data can be understood as the map information of the target scenario, which may include information such as the layout of the road, traffic signs, and obstacle positions. The historical trajectory data may be the driving trajectory information of each intelligent vehicle in the target scenario during a past period (historical moment), and generally may include information such as the position, speed, and acceleration of the vehicle. The current driving state may be the driving situation of the vehicle at the current moment, such as the driving position, speed, acceleration, etc.

[0071] The candidate prediction trajectory can be understood as multiple driving trajectories of the target vehicle predicted at a future moment based on the historical trajectory data and scenario map data of the vehicle. The trajectory prediction probability value is the probability of each candidate prediction trajectory occurring. Each candidate prediction trajectory corresponds to a trajectory prediction probability value. The risk cost value may be the risk cost that may occur when evaluating the target vehicle driving according to a certain candidate prediction trajectory when considering risk allocation in a traffic scenario. For example, the risk of collision between the target vehicle and other vehicles, etc. By considering the risk cost, the candidate prediction trajectory corresponding to the highest trajectory prediction probability value is optimized, and the final predicted trajectory of the target vehicle at a future moment is the target trajectory.

[0072] Specifically, in this embodiment, first, obtain the map data of the scene where the target vehicle is located, the historical trajectory data of multiple intelligent vehicles (the target vehicle and surrounding vehicles) in the target scenario at the historical moment, and the current driving state of the target vehicle. After obtaining this data, use the historical trajectory data and scenario map data of the intelligent vehicle to predict multiple possible candidate prediction trajectories of the target vehicle at a future moment, and at the same time, calculate the trajectory prediction probability value of each candidate prediction trajectory. Then, according to the trajectory prediction probability value of each candidate prediction trajectory and the current driving state of each intelligent vehicle, evaluate the risk cost that the target vehicle may face when driving according to each candidate prediction trajectory, so as to obtain the risk cost value of the target vehicle. Finally, among all candidate prediction trajectories, optimize the trajectory with the highest prediction probability value through the risk cost value, so as to generate the target trajectory of the target vehicle at a future moment.

[0073] It is understandable that the embodiments of the present application provide a trajectory prediction method for an autonomous vehicle. The method includes: obtaining scene map data of a target scene where a target vehicle is located, historical trajectory data of multiple intelligent vehicles in the target scene, and the current driving state of the target vehicle, where the intelligent vehicles include the target vehicle and multiple surrounding vehicles; determining multiple candidate prediction trajectories of the target vehicle at a future moment and trajectory prediction probability values corresponding to the multiple candidate prediction trajectories according to the historical trajectory data of the multiple intelligent vehicles and the scene map data; calculating a risk cost value of the target vehicle according to the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories of the target vehicle at the future moment and the current driving states of the multiple intelligent vehicles; and optimizing the candidate prediction trajectory with the highest trajectory prediction probability value according to the risk cost value of the target vehicle to generate a target trajectory of the target vehicle at the future moment. The present application generates multiple candidate prediction trajectories of the target vehicle through the historical trajectory data of the intelligent vehicles and the target scene map data, and optimizes the candidate prediction trajectories through the calculated risk cost value of the target vehicle to improve the safety and adaptability of the trajectory prediction of the intelligent vehicle in a complex traffic scene.

[0074] In a possible implementation manner, determining multiple candidate prediction trajectories of the target vehicle at a future moment and trajectory prediction probability values corresponding to the multiple candidate prediction trajectories according to the historical trajectory data of the multiple intelligent vehicles and the scene map data includes:

[0075] Encoding the historical trajectory data of the multiple intelligent vehicles through a trajectory encoder to obtain encoded historical trajectory data of the multiple intelligent vehicles.

[0076] Obtaining relative position data between the multiple intelligent vehicles according to the encoded historical trajectory data of the multiple intelligent vehicles.

[0077] Determining interaction features of the multiple intelligent vehicles according to the scene map data, the encoded historical trajectory data of the multiple intelligent vehicles, and the relative position data between the multiple intelligent vehicles.

[0078] Extracting features from the encoded historical trajectory data through a multi-layer perceptron to determine intention features of the multiple intelligent vehicles and intention prediction probability values corresponding to the intention features, where the intention features include lateral intention features and longitudinal intention features.

[0079] Obtaining multiple candidate prediction trajectories of the target vehicle at a future moment and trajectory prediction probability values corresponding to the multiple candidate prediction trajectories according to the interaction features of the multiple intelligent vehicles and the intention features of the multiple intelligent vehicles.

[0080] In this embodiment, joint prediction is implemented through a feature extraction module based on the deep learning architecture Transformer, that is, the interaction features are extracted by using the context of the target scenario and the driving information of the intelligent vehicle.

[0081] The trajectory encoder can be a model for encoding the historical trajectory data of the intelligent vehicle. The historical trajectory data of the intelligent vehicle is encoded by the trajectory encoder to obtain the encoded historical trajectory data. The encoded historical trajectory data can contain the key features of the trajectory, usually represented in the form of vectors or matrices.

[0082] The relative position data can be understood as the information describing the relative positions between each intelligent vehicle. For example, the target vehicle is on the left, right or front of the surrounding vehicles, which is conducive to the understanding of the spatial relationship between vehicles.

[0083] The interaction features can be the features reflecting the interaction or influence between multiple intelligent vehicles, or can also reflect the features between the intelligent vehicle and the scenario. The interaction features can include features such as the distance, speed difference, and acceleration difference between the target vehicle and the surrounding vehicles, as well as the distance, speed difference, and acceleration difference between the surrounding vehicles, and can also include features such as the distance between the vehicle and the road edge and the relative position of the vehicle and the traffic sign.

[0084] A multi-layer perceptron (MLP) is a feedforward neural network that can consist of multiple layers, with multiple neurons in each layer, capable of feature extraction and pattern recognition for input data. In this embodiment, by using a multi-layer perception layer, feature extraction can be performed on the historical trajectory data of all intelligent vehicles, thereby obtaining the intention features of all intelligent vehicles. The intention features can be understood as features used to describe the driving intentions of intelligent vehicles. Due to the complexity and diversity of driving operations, the actual trajectories of vehicles in real traffic scenarios are usually uncertain. It is considered that the main lateral driving operations of intelligent vehicles are left turn (LT), straight (ST), and right turn (RT), and the main longitudinal driving operations are acceleration (ACC), constant speed (CON), and deceleration (DEC). Therefore, the obtained intention features of all intelligent vehicles can include lateral intention features (left turn, straight, right turn) and longitudinal intention features (acceleration, constant speed, deceleration). After extracting the intention features of intelligent vehicles, the MLP with a softmax activation function calculates the intention prediction probability values corresponding to the intention features according to the intention features. The intention prediction probability values can reflect the probability of each intention feature occurring, indicating the probability of an intelligent vehicle executing a certain driving intention. The intention prediction probability values can include la prediction probability values (intention prediction probability values corresponding to lateral intention features) and lo prediction probability values (intention prediction probability values corresponding to longitudinal intention features), where la ∈ {LT, ST, RT} and lo ∈ {ACC, DEC, CON}.

[0085] In a specific embodiment, first, a trajectory encoder is used to encode the historical trajectory data of multiple intelligent vehicles to obtain the encoded historical trajectory data. After obtaining the encoded historical trajectory data, the relative positions between multiple intelligent vehicles are calculated based on the encoded historical trajectory data, thereby understanding the spatial relationship between the vehicles. Then, combining the scenario map data, the encoded historical trajectory data, and the relative position data, the interaction features between multiple intelligent vehicles are determined. These interaction features reflect the mutual influence between vehicles and the relationship between vehicles and roads, providing feature support for predicting the future trajectories of vehicles. Then, a multi-layer perceptron is used to extract features from the encoded historical trajectory data to determine the intention features of multiple intelligent vehicles and the intention prediction probability values corresponding to each intention feature. The intention features and the intention prediction probability values reflect the driving intentions of intelligent vehicles and the possibility of each intention occurring, and the uncertainty of trajectory prediction can be reduced through intention recognition. Finally, based on the extracted interaction features and intention features of multiple intelligent vehicles, multiple candidate prediction trajectories of the target vehicle at a future moment and the trajectory prediction probability values corresponding to each candidate prediction trajectory are generated.

[0086] It should be understood that through the encoding process of historical trajectory data, the calculation of the relative positions between intelligent vehicles, and the extraction of interaction features and intention features, more accurate feature information is provided for the trajectory prediction of the target vehicle at future moments.

[0087] It should be noted that before encoding the historical trajectory data of intelligent vehicles, the historical trajectory data is first subjected to coordinate system conversion (i.e., converting the local relative coordinate system into the global coordinate system) and normalization processing to facilitate subsequent encoding of the historical trajectory data.

[0088] In a possible implementation manner, based on the scenario map data, the encoded historical trajectory data of multiple intelligent vehicles, and the relative position data between multiple intelligent vehicles, the interaction features of multiple intelligent vehicles are determined, including:

[0089] The scenario map data is encoded through a map encoder to obtain the encoded scenario map data.

[0090] Based on the encoded historical trajectory data of multiple intelligent vehicles, the first interaction features between multiple intelligent vehicles are obtained.

[0091] Based on the encoded historical trajectory data of multiple intelligent vehicles and the encoded scenario map data, the second interaction features between multiple intelligent vehicles and the map are obtained.

[0092] Based on the first interaction features, the second interaction features, and the relative position data between multiple intelligent vehicles, the interaction features of multiple intelligent vehicles are obtained.

[0093] The map encoder is a model used to encode the scenario map data and can convert the initial scenario map data into forms such as vectors or matrices. The encoded scenario map data can be understood as the scenario map data obtained after the initial scenario map data is processed by the map encoder, and the encoded scenario map data can contain key features of the target scenario map, such as map boundaries, road directions, and other information.

[0094] The first interaction feature can be understood as features such as the distance, speed difference, and acceleration difference between multiple intelligent vehicles. The first interaction feature is an interaction feature extracted based on the encoded historical trajectory data of multiple intelligent vehicles. The second interaction feature can be understood as the features between multiple intelligent vehicles and the scene map, which can include the distance between the vehicle and the road edge, the relative position of the vehicle and the traffic sign, etc. The second interaction feature is an interaction feature extracted based on the encoded historical trajectory data of multiple intelligent vehicles and the encoded scene map data. The interaction feature is the comprehensive feature between vehicles obtained by comprehensively considering the first interaction feature, the second interaction feature, and the relative position data between multiple intelligent vehicles. The interaction feature can more comprehensively describe the interaction and influence between multiple intelligent vehicles.

[0095] In a specific embodiment, first, the scene map data is encoded by a map encoder to obtain the encoded scene map data; then, this encoded scene map data can be combined with the encoded historical trajectory data of multiple intelligent vehicles to extract the second interaction feature between the vehicle and the map. At the same time, the first interaction feature between vehicles is extracted from the encoded historical trajectory data of multiple intelligent vehicles; then, by comprehensively considering the first interaction feature, the second interaction feature, and the relative position data between multiple intelligent vehicles, the interaction feature of multiple intelligent vehicles is obtained.

[0096] It should be understood that in this embodiment, these interaction features can more comprehensively describe the interaction and influence between vehicles and between vehicles and the map, providing important feature information for the trajectory prediction of autonomous vehicles.

[0097] In a possible implementation manner, according to the interaction features of multiple intelligent vehicles and the intention features of multiple intelligent vehicles, multiple candidate prediction trajectories of the target vehicle at a future moment and the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories are obtained, including:

[0098] Fuse the interaction features of multiple intelligent vehicles and the intention features of multiple intelligent vehicles to obtain a fused feature.

[0099] Concatenate the fused feature and the intention prediction probability value corresponding to the intention feature, and input them into a multi-layer perceptron MLP for processing to obtain the embedding vector corresponding to the fused feature.

[0100] Input the embedding vector into a long short-term memory LSTM decoder to obtain multiple candidate prediction trajectories of the target vehicle at a future moment and the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories.

[0101] In a specific embodiment, after obtaining the interaction features and intention features of multiple intelligent vehicles, the interaction features and intention features (lateral features and longitudinal features) of the multiple intelligent vehicles can be fused first to obtain the fusion features of the multiple intelligent vehicles. The fusion features aggregate the feature information of all intelligent vehicles. The fusion process can be weighted, concatenated, convolved, or other feature combination methods, which can more comprehensively describe the interaction between vehicles and the driving intention of the vehicles. After obtaining the fusion features of the multiple intelligent vehicles, the fusion features and the intention prediction probability values corresponding to the intention features are concatenated to form a feature vector, and the feature vector is input into a multi-layer perceptron (MLP) for processing to obtain the embedding vector corresponding to the fusion features. The embedding vector can be a vector representation obtained by processing the fusion features and the intention prediction probability values through a multi-layer perceptron, which can capture the complex patterns and relationships of the data.

[0102] After obtaining the embedding vector corresponding to the fusion features, the embedding vector is input into a Long Short-Term Memory (LSTM) decoder. The LSTM decoder uses the information in the embedding vector to generate multiple candidate prediction trajectories of the target vehicle at future times. At the same time, the estimated prediction probability values corresponding to the multiple candidate prediction trajectories are obtained. Among them, the long short-term memory LSTM decoder is a special recurrent neural network structure that can learn long-term dependencies and is suitable for the decoding task of sequence data.

[0103] It should be understood that in this embodiment, the interaction between vehicles and the driving intention of the vehicles are comprehensively considered, and the learning ability of the neural network is used to provide data support for the trajectory prediction of autonomous vehicles.

[0104] In a possible implementation manner, according to the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories of the target vehicle at future times and the current driving states of the multiple intelligent vehicles, the risk cost value of the target vehicle is calculated, including:

[0105] According to the current driving states of the multiple intelligent vehicles, the collision probability values and estimated injury values of the multiple intelligent vehicles on the candidate prediction trajectories with the highest trajectory prediction probability values are determined.

[0106] According to the collision probability values and the estimated injury values, the risk values of each intelligent vehicle for collision are obtained.

[0107] According to the risk values of each intelligent vehicle for collision, the risk cost value of the target vehicle is calculated.

[0108] The collision probability value can be understood as the probability value of a smart vehicle colliding with other smart vehicles, which is usually calculated based on the current driving state of the smart vehicle (such as the vehicle's speed, acceleration, and relative position between vehicles, etc.). The estimated injury value can be understood as the degree of injury expected to be caused in the event of a collision. The estimated injury value can include the degree of vehicle damage, the degree of passenger injury, etc., and is usually estimated based on the situation of the collision. The risk value can be the risk degree of each smart vehicle colliding with the target vehicle, which can be calculated by combining the collision probability value and the estimated injury value. Generally, the higher the risk value, the greater the likelihood of a collision. The risk cost value can be the total risk cost obtained by comprehensively considering the risk values of all smart vehicles when the target vehicle adopts a certain candidate prediction trajectory at a future moment. The risk cost value can be used to evaluate the safety and feasibility of different trajectories.

[0109] Specifically, based on the current heading angle and size (length and width of the vehicle body) of the vehicle, the front-end position and rear-end position of the vehicle can be calculated, and the front-end position, rear-end position, and the center position of the vehicle are used for collision detection together. At this time, the collision probability value is defined as conforming to a multivariate normal distribution. By calculating the normal distributions (i.e., collision probability values) of the center, front-end, and rear-end positions of the vehicle and superimposing the normal distributions of the three positions, the collision probability value of the vehicle at the current moment is obtained. For objectivity, this application only considers factors that do not change due to human intervention, such as the mass, speed, and deflection angle of both colliding vehicles. According to the angle of the collision, the collision type can be classified as a front collision, a side collision, or a rear collision. And, to simplify the collision model, the collision area is symmetrically processed, and the collision model is divided into protected (vehicles, trucks, etc.) road users and unprotected (pedestrians, cyclists, etc.) road users.

[0110] At the same time, to calculate the estimated injury value, damage calculation equations are introduced, as shown in formulas (1) and (2).

[0111]

[0112] In the above formulas (1) and (2), H is the estimated injury value of road user A after the collision between road user A and road user B, μ0, μ1, and μ area are coefficients determined according to experience (i.e., constant values, which are obtained by fitting with real data), Δv A is the relative speed difference of road user A relative to road user B, m A is the mass of road user A, m B is the mass of road user B, v A is the speed of road user A, v Bis the speed of road user B, and α is the collision angle between road user A and road user B.

[0113] Using the damage calculation equation, an estimated injury value is assigned to the collision probability value of each intelligent vehicle. For the convenience of calculation, a risk value is used to describe each possible candidate prediction trajectory. Therefore, the maximum risk in the future moment is selected as the risk value of each candidate prediction trajectory. Here, only the risk value of the possible collision between the target vehicle and the surrounding vehicles is calculated, as shown in formula (3).

[0114] R = max t (H·P t ) (3)

[0115] In formula (3), R represents the risk value of the vehicle, H is the estimated injury value of road user A after the collision between road user A and road user B, and P t is the collision probability value of the candidate prediction trajectory with the highest trajectory prediction probability value.

[0116] After obtaining the risk value, the risk allocation concept is reflected in the cost function, considering different aspects of driving decisions, including the self - protection risk cost value, the care cost value for vulnerable groups, and the sudden high - risk cost value for attention to prominent high risks. Further, according to the risk value of each intelligent vehicle's collision, the risk cost value of the target vehicle is calculated, including at least one of the following:

[0117] According to the risk value of each intelligent vehicle and the preset boundary damage value on the candidate prediction trajectory with the highest trajectory prediction probability value, the self - protection risk cost value of the target vehicle is calculated.

[0118] According to the difference between the risk values of any two intelligent vehicles on the candidate prediction trajectory with the highest trajectory prediction probability value, the care cost value of the target vehicle is calculated.

[0119] According to the risk value of each intelligent vehicle and the preset scaling factor on the candidate prediction trajectory with the highest trajectory prediction probability value, the sudden high - risk cost value of the target vehicle is calculated.

[0120] According to the self - protection risk cost value, care cost value and sudden high - risk cost value of the target vehicle, the risk cost value of the target vehicle is calculated.

[0121] Specifically, first, considering self - protection on the candidate prediction trajectory, the self - protection risk cost value of the target vehicle on the candidate prediction trajectory is calculated according to the following formula (4), that is:

[0122]

[0123] Among them, cs is the self - protection cost value, R i is the risk value of each road user calculated according to formula (3), where i represents the n road users in the target scenario, and R b is the preset boundary damage value. The preset boundary damage value represents the risk value of the intelligent vehicle colliding with the road, which is determined based on experience and is triggered when the vehicle's trajectory in the future has an intersection with the road boundary. The self - protection risk cost value reflects the cost of the target vehicle's self - protection to avoid damage.

[0124] Accumulate and normalize the risks of all detected intelligent vehicles in the scenario. Select the trajectory with the lowest risk cost value according to the Bayesian principle, and ensure that the target vehicle makes the best trajectory decision based on this principle.

[0125] After that, consider the vulnerable groups on the candidate prediction trajectory, and calculate the care cost value of the target vehicle on the candidate prediction trajectory according to the following formula (5), that is:

[0126]

[0127] In formula (5), c c is the care cost value, representing the average difference between different risk values. R i and R j are the risk values of any two road users. It should be understood that if the difference in risk values between the protected road user and the unprotected road user is greater, it indicates a higher care cost value, aiming to avoid causing a disproportionately high risk to the vulnerable groups in the process of pursuing a lower safety cost.

[0128] Then, consider the sudden high risk on the candidate prediction trajectory, and calculate the sudden high - risk cost value of the target vehicle on the candidate prediction trajectory according to the following formula (6), that is:

[0129]

[0130] In formula (6), c r is the sudden high - risk cost value, and R i is the risk value of each road user calculated according to formula (3). When dealing with sudden high risks, formula (6) can evaluate the risk cost of the candidate prediction trajectory according to the maximin principle to ensure that the vehicle's performance is still acceptable in the most unfavorable situation (i.e., higher risk). Evaluate the potential risk by calculating the maximin value between self - protection and obstacles, and use the scaling factor f(R i) To adjust the final risk cost can ensure the rationality of the final result. It should be understood that in the face of sudden high risks, the situations that may cause the greatest harm will be given priority to ensure that the risk cost can be effectively controlled even under the worst conditions.

[0131] Finally, the risk cost value of the target vehicle can be as shown in formula (7).

[0132] L risk = ω s ·c s + ω c ·c c + ω r ·c r (7)

[0133] In formula (7), L risk is the risk cost value, c s is the self - protection cost value, c c is the care cost value, c r is the sudden high - risk cost value, ω s is the weight corresponding to the self - protection cost value, ω c is the weight corresponding to the care cost value, ω r is the weight corresponding to the sudden high - risk cost value.

[0134] It should be understood that the risk cost value can be used to evaluate the overall risk level of the predicted trajectory and provide a reference for the decision - making of the autonomous vehicle.

[0135] In a possible implementation manner, the trajectory prediction method of the autonomous vehicle further includes:

[0136] Based on a preset loss function, calculate the loss value between the target trajectory and the candidate prediction trajectory, where the loss value includes the smooth loss value of the displacement deviation and the cross - entropy loss value of the intention recognition.

[0137] According to the risk cost value, the smooth loss value of the displacement deviation, and the cross - entropy loss value of the intention recognition, optimize the risk optimization module to obtain an optimized risk optimization module, where the risk optimization module is used to optimize the candidate prediction trajectory with the highest trajectory prediction probability value.

[0138] In this embodiment, the risk optimization module is a module that optimizes the candidate prediction trajectory with the highest trajectory prediction probability value. It considers the smooth loss value of the displacement deviation and the cross - entropy loss value of the intention recognition, and introduces the risk cost value to optimize the candidate prediction trajectory of the target vehicle. The candidate prediction trajectory is adjusted through an optimization algorithm to make it optimal in terms of safety and accuracy. When optimizing through the risk optimization module, it is necessary to perform data training on the risk optimization module to obtain an optimized risk optimization module.

[0139] In this embodiment, the preset loss function can be a predefined mathematical function that can measure the difference between the target trajectory and the candidate prediction trajectory. Usually, the optimal parameters are obtained by minimizing the loss value. The loss value is the error representing the difference between the target trajectory and the candidate prediction trajectory calculated according to the preset loss function. In this embodiment, the loss value includes the smooth loss value of the displacement deviation and the cross-entropy loss value of the intention recognition. Among them, the smooth loss value of the displacement deviation measures the deviation of the candidate prediction trajectory from the target trajectory in terms of displacement, and is usually used to ensure the continuity and smoothness of the prediction trajectory and avoid sharp changes. The cross-entropy loss value of the intention recognition measures the difference between the driving intention reflected by the candidate prediction trajectory and the driving intention reflected by the target trajectory. By minimizing the cross-entropy loss value of the intention recognition, the accuracy of the trajectory prediction can be improved, and the prediction trajectory can better reflect the driver's intention.

[0140] Specifically, in this embodiment, the model is trained by a state-based strategy. In the first five rounds of training, the preset loss function is used to evaluate the deviation between the predicted candidate prediction trajectory and the target trajectory, which can include two parts: the smooth loss value of the displacement deviation and the cross-entropy loss value of the intention recognition. Among them, the smooth loss value of the displacement deviation L pre can be calculated by formula (8).

[0141]

[0142] In formula (8), L pre is the smooth loss value of the displacement deviation, is the predicted trajectory of the current training sample, y i is the target trajectory of the current training sample, and i represents the n intelligent vehicles in the target scenario.

[0143] The cross-entropy loss value of the intention recognition L man can be calculated by formula (9).

[0144] L man = -∑ la,lo Q gt ·logQ pre (9)

[0145] Among them, L man is the cross-entropy loss value of the intention recognition, la is the lateral intention feature, lo is the longitudinal intention feature, Q gt is the true intention of the current training sample, and Q pre is the predicted intention of the current training sample.

[0146] Therefore, the strategy for the first five rounds is shown in formula (10):

[0147] L1 = L pre + τL man (10)

[0148] In Equation (10), L1 is the loss value of the first five rounds, and τ is the loss value weight.

[0149] In subsequent model training, the impact of the risk cost value on the generation of the predicted trajectory is considered. Therefore, the strategy after five rounds is as shown in Equation (11):

[0150] L2 = L pre + τL man +(1 - τ)L risk (11)

[0151] In Equation (11), L2 is the loss value after five rounds, and τ is the loss value weight.

[0152] It should be understood that in this embodiment, by making the loss value L2 the minimum value, each parameter in the risk optimization module is optimized to obtain an optimized risk optimization module, thereby improving the accuracy and safety of the trajectory prediction of the autonomous driving vehicle.

[0153] Exemplarily, as Figure 3 shown Figure 3 is a schematic flowchart of a trajectory prediction method for an autonomous driving vehicle provided in another embodiment of the present application. Figure 3 In it, the historical trajectory data of multiple intelligent vehicles in the target scenario and the scenario map data of the target scenario are input into the interaction module. In the interaction module, the historical trajectory data is encoded by the trajectory encoder to obtain the relative position data (i.e., r i→j = [α i→j , β i→j , d i→j ) between multiple intelligent vehicles and the interaction feature between intelligent vehicles (i.e., agent-agent interaction feature), and the interaction feature between the intelligent vehicle and the map (i.e., agent-map interaction feature) is obtained by encoding the scenario map data by the map encoder and combining the interaction feature between intelligent vehicles. Then, the agent-agent interaction feature, the agent-map interaction feature, and the relative position data are fused to obtain a fusion feature, which can help generate a reasonable trajectory that conforms to the actual situation in the future. Then, the lateral intention feature la and the longitudinal intention feature lo are extracted from the encoded historical trajectory data through the intention module. Then, the lateral intention feature la and the longitudinal intention feature lo are concatenated with the above fusion feature and input into the LSTM decoder together to obtain multiple candidate predicted trajectories of the target vehicle at future moments and the trajectory prediction probability values corresponding to the multiple candidate predicted trajectories; finally, the risk optimization module performs risk optimization on the candidate predicted trajectories and outputs the final target trajectory. Figure 2where α i→j = [α i→j , β i→j , d i→j is used to represent the relative position data between different intelligent vehicles i and j. α i→j is the heading difference vector, β i→j is the relative azimuth angle vector, d i→j is the displacement difference vector, and Q, V, and K are the query Q, key K, and value V in the Transformer architecture.

[0154] It should be understood that in this embodiment, the trajectory prediction is enhanced by focusing on the trajectories related to the intention rather than considering all possible motion patterns, thereby improving the accuracy and efficiency of the autonomous driving system. At the same time, by introducing a risk optimization module, potential risks are considered for the predicted trajectories, enabling the vehicle to make more adaptable and safer decisions.

[0155] It should be understood that the magnitudes of the sequence numbers of the steps in the above embodiments do not mean the order of execution. The execution order of each process should be determined according to its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application.

[0156] Corresponding to a trajectory prediction method for an autonomous driving vehicle in the above embodiment, Figure 4 FIG. shows a schematic structural diagram of a trajectory prediction device for an autonomous driving vehicle provided in an embodiment of the present application. For the sake of convenience of description, only the parts related to the embodiments of the present application are shown.

[0157] Referring to Figure 4 , the trajectory prediction device 3 of the autonomous driving vehicle in this embodiment includes:

[0158] An acquisition module 31, configured to acquire the scene map data of the target scene where the target vehicle is located, the historical trajectory data of multiple intelligent vehicles in the target scene, and the current driving states of the multiple intelligent vehicles, where the intelligent vehicles include the target vehicle and multiple surrounding vehicles.

[0159] A prediction module 32, configured to determine multiple candidate prediction trajectories of the target vehicle at a future moment and the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories according to the historical trajectory data of the multiple intelligent vehicles and the scene map data.

[0160] A calculation module 33, configured to calculate the risk cost value of the target vehicle according to the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories of the target vehicle at a future moment and the current driving states of the multiple intelligent vehicles.

[0161] ​

[0162] It can be understood that this embodiment provides a trajectory prediction device for an autonomous vehicle. The trajectory prediction device 3 of the autonomous vehicle obtains the scene map data of the target scene where the target vehicle is located, the historical trajectory data of multiple intelligent vehicles in the target scene, and the current driving state of the target vehicle through the acquisition module 31. Among them, the intelligent vehicles include the target vehicle and multiple surrounding vehicles; the prediction module 32 determines multiple candidate prediction trajectories of the target vehicle at future moments and the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories according to the historical trajectory data and scene map data of the multiple intelligent vehicles; the calculation module 33 calculates the risk cost value of the target vehicle according to the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories of the target vehicle at future moments and the current driving states of the multiple intelligent vehicles; the optimization module 34 optimizes the candidate prediction trajectory with the highest trajectory prediction probability value according to the risk cost value of the target vehicle to generate the target trajectory of the target vehicle at future moments. This device generates multiple candidate prediction trajectories of the target vehicle through the historical trajectory data of the intelligent vehicles and the target scene map data, and optimizes the candidate prediction trajectories through the calculated risk cost value of the target vehicle to improve the safety and adaptability of the trajectory prediction of the intelligent vehicle in complex traffic scenarios.

[0163] Optionally, the prediction module 32 includes:

[0164] The first encoding processing sub-module is used to encode the historical trajectory data of multiple intelligent vehicles through a trajectory encoder to obtain the encoded historical trajectory data of the multiple intelligent vehicles.

[0165] The relative position extraction sub-module is used to obtain the relative position data between multiple intelligent vehicles according to the encoded historical trajectory data of the multiple intelligent vehicles.

[0166] The interaction feature determination sub-module is used to determine the interaction features of multiple intelligent vehicles according to the scene map data, the encoded historical trajectory data of multiple intelligent vehicles, and the relative position data between multiple intelligent vehicles.

[0167] The intention feature extraction sub-module is used to extract features from the encoded historical trajectory data through a multi-layer perceptron to determine the intention features of multiple intelligent vehicles and the intention prediction probability values corresponding to the intention features, where the intention features include lateral intention features and longitudinal intention features.

[0168] The candidate trajectory prediction sub-module is used to obtain multiple candidate prediction trajectories of the target vehicle at future moments and the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories according to the interaction features of multiple intelligent vehicles and the intention features of multiple intelligent vehicles.

[0169] Optionally, the candidate trajectory prediction sub-module includes:

[0170] A feature fusion unit for fusing the interaction features of multiple intelligent vehicles and the intention features of multiple intelligent vehicles to obtain fused features.

[0171] A feature concatenation unit for concatenating the fused features and the intention prediction probability values corresponding to the intention features and inputting them into a multi-layer perceptron (MLP) for processing to obtain the embedding vector corresponding to the fused features.

[0172] A trajectory prediction unit for inputting the embedding vector into a long short-term memory (LSTM) decoder to obtain multiple candidate prediction trajectories of the target vehicle at future moments and the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories.

[0173] Optionally, the interaction feature determination sub-module includes:

[0174] A second encoding processing unit for encoding the scene map data through a map encoder to obtain the encoded scene map data.

[0175] A first interaction feature determination unit for obtaining the first interaction features between multiple intelligent vehicles based on the encoded historical trajectory data of the multiple intelligent vehicles.

[0176] A second interaction feature determination unit for obtaining the second interaction features between multiple intelligent vehicles and the map based on the encoded historical trajectory data of the multiple intelligent vehicles and the encoded scene map data.

[0177] An interaction feature determination unit for obtaining the interaction features of multiple intelligent vehicles based on the first interaction features, the second interaction features, and the relative position data between the multiple intelligent vehicles.

[0178] Optionally, the calculation module 33 includes:

[0179] A collision probability value determination sub-unit for determining the collision probability value and the estimated injury value of multiple intelligent vehicles on the candidate prediction trajectory with the highest trajectory prediction probability value according to the current driving states of the multiple intelligent vehicles.

[0180] A risk value determination sub-unit for obtaining the risk value of each intelligent vehicle for collision according to the collision probability value and the estimated injury value.

[0181] A risk cost value determination sub-unit for calculating the risk cost value of the target vehicle according to the risk value of each intelligent vehicle for collision.

[0182] Optionally, the risk cost value determination sub-unit includes at least one of the following:

[0183] The first risk determination unit is configured to calculate the self - protection risk cost value of the target vehicle according to the risk values of each intelligent vehicle on the candidate prediction trajectory with the highest trajectory prediction probability value and a preset boundary damage value.

[0184] The second risk determination unit is configured to calculate the care cost value of the target vehicle according to the difference between the risk values of any two intelligent vehicles on the candidate prediction trajectory with the highest trajectory prediction probability value.

[0185] The third risk determination unit is configured to calculate the sudden high - risk cost value of the target vehicle according to the risk values of each intelligent vehicle on the candidate prediction trajectory with the highest trajectory prediction probability value and a preset scaling factor.

[0186] The risk cost value determination unit is configured to calculate the risk cost value of the target vehicle according to the self - protection risk cost value, the care cost value, and the sudden high - risk cost value of the target vehicle.

[0187] Optionally, the trajectory prediction device 3 of the autonomous vehicle further includes:

[0188] The loss value determination module is configured to calculate the loss value between the target trajectory and the candidate prediction trajectory based on a preset loss function, where the loss value includes the smooth loss value of the displacement deviation and the cross - entropy loss value of the intention recognition.

[0189] The optimization module is configured to optimize the risk optimization module according to the risk cost value, the smooth loss value of the displacement deviation, and the cross - entropy loss value of the intention recognition, so as to obtain an optimized risk optimization module, where the risk optimization module is configured to optimize the candidate prediction trajectory with the highest trajectory prediction probability value.

[0190] It should be noted that for the information interaction, execution process, etc. among the modules in the above - mentioned trajectory prediction device 3 of the autonomous vehicle, since they are based on the same concept as the method embodiment of the present application, for their specific functions and the technical effects brought, reference can be made to the method embodiment part, and details are not described here again.

[0191] The embodiment of the present application also provides a terminal device, as Figure 5 shown, Figure 5 is a schematic structural diagram of a terminal device provided by an embodiment of the present application. Referring to Figure 5 , the terminal device 4 in this embodiment includes: a memory 41, a processor 42, and a computer program stored in the memory 41 and executable on the processor 42. When the processor 42 executes the computer program, it implements the steps in any of the above - mentioned method embodiments of the trajectory prediction method for the autonomous vehicle.

[0192] The embodiments of the present application further provide a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in the above-mentioned method embodiments can be implemented.

[0193] The embodiments of the present application provide a computer program product. When the computer program product runs on a mobile terminal, the mobile terminal is enabled to implement the steps in the above-mentioned method embodiments when executed.

[0194] 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, to implement all or part of the processes in the above-mentioned method embodiments of the present application, a computer program can be used to instruct the relevant hardware to complete. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file or some intermediate form, etc. The computer-readable medium can at least include: any entity or device that can carry the computer program code to the photographing device / terminal device, recording medium, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal, and software distribution medium. For example, a USB flash drive, a mobile hard disk, a magnetic disk or an optical disc, etc. In some jurisdictions, according to legislation and patent practice, the computer-readable medium may not be an electrical carrier signal and a telecommunication signal.

[0195] In the above embodiments, the descriptions of the respective embodiments have their own focuses. For parts not detailed or recorded in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0196] Those of ordinary skill in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

[0197] In the embodiments provided in the present application, it should be understood that the disclosed device / network device and method can be implemented in other ways. For example, the device / network device embodiments described above are merely illustrative. For example, the division of modules or units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple 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 displayed or discussed coupling or direct coupling or communication connection between each other can be through some interfaces. The indirect coupling or communication connection of the device or unit can be in electrical, mechanical or other forms.

[0198] The units described as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0199] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A trajectory prediction method for an autonomous vehicle, characterized in that, Including: Obtain the scene map data of the target scene where the target vehicle is located, the historical trajectory data of multiple intelligent vehicles in the target scene, and the current driving states of the multiple intelligent vehicles, where the intelligent vehicles include the target vehicle and multiple surrounding vehicles; According to the historical trajectory data and the scene map data of the multiple intelligent vehicles, determine multiple candidate prediction trajectories of the target vehicle at a future moment and the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories; According to the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories of the target vehicle at a future moment and the current driving states of the multiple intelligent vehicles, calculate the risk cost value of the target vehicle; According to the risk cost value of the target vehicle, optimize the candidate prediction trajectory with the highest trajectory prediction probability value to generate the target trajectory of the target vehicle at a future moment.

2. The trajectory prediction method for an autonomous vehicle according to claim 1, characterized in that, The step of determining multiple candidate prediction trajectories of the target vehicle at a future moment and the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories according to the historical trajectory data and the scene map data of the multiple intelligent vehicles includes: Encode the historical trajectory data of the multiple intelligent vehicles through a trajectory encoder to obtain the encoded historical trajectory data of the multiple intelligent vehicles; According to the encoded historical trajectory data of the multiple intelligent vehicles, obtain the relative position data between the multiple intelligent vehicles; According to the scene map data, the encoded historical trajectory data of the multiple intelligent vehicles, and the relative position data between the multiple intelligent vehicles, determine the interaction features of the multiple intelligent vehicles; Extract features from the encoded historical trajectory data through a multi-layer perceptron to determine the intention features of the multiple intelligent vehicles and the intention prediction probability values corresponding to the intention features, where the intention features include lateral intention features and longitudinal intention features; According to the interaction features of the multiple intelligent vehicles and the intention features of the multiple intelligent vehicles, obtain the multiple candidate prediction trajectories of the target vehicle at a future moment and the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories.

3. The trajectory prediction method for an autonomous vehicle according to claim 2, wherein, The step of obtaining the multiple candidate prediction trajectories of the target vehicle at a future moment and the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories according to the interaction features of the multiple intelligent vehicles and the intention features of the multiple intelligent vehicles includes: Fuse the interaction features of the multiple intelligent vehicles and the intention features of the multiple intelligent vehicles to obtain a fused feature; Concatenate the fused feature and the intention prediction probability value corresponding to the intention feature, and input them into a multi-layer perceptron MLP for processing to obtain the embedding vector corresponding to the fused feature; Input the embedding vector into a long short-term memory LSTM decoder to obtain the multiple candidate prediction trajectories of the target vehicle at a future moment and the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories.

4. The trajectory prediction method for an autonomous vehicle according to claim 2, wherein Determining interaction features of multiple intelligent vehicles based on the scenario map data, the encoded historical trajectory data of multiple intelligent vehicles, and the relative position data between multiple intelligent vehicles includes: Encoding the scenario map data through a map encoder to obtain encoded scenario map data; Obtaining first interaction features between multiple intelligent vehicles based on the encoded historical trajectory data of multiple intelligent vehicles; Obtaining second interaction features between multiple intelligent vehicles and the map based on the encoded historical trajectory data of multiple intelligent vehicles and the encoded scenario map data; Obtaining interaction features of multiple intelligent vehicles based on the first interaction features, the second interaction features, and the relative position data between multiple intelligent vehicles.

5. The trajectory prediction method for an autonomous vehicle according to any one of claims 1-4, characterized in that, Calculating a risk cost value of the target vehicle based on the trajectory prediction probability values corresponding to multiple candidate prediction trajectories of the target vehicle at a future moment and the current driving states of multiple intelligent vehicles includes: Determining collision probability values and estimated injury values of multiple intelligent vehicles on the candidate prediction trajectory with the highest trajectory prediction probability value according to the current driving states of multiple intelligent vehicles; Obtaining a risk value of collision for each intelligent vehicle according to the collision probability value and the estimated injury value; Calculating the risk cost value of the target vehicle according to the risk values of collision for each intelligent vehicle.

6. The trajectory prediction method for an autonomous vehicle according to claim 5, wherein, Calculating the risk cost value of the target vehicle according to the risk values of collision for each intelligent vehicle includes at least one of the following: Calculating a self - protection risk cost value of the target vehicle according to the risk value of each intelligent vehicle and a preset boundary damage value on the candidate prediction trajectory with the highest trajectory prediction probability value; Calculating a care cost value of the target vehicle according to the difference between the risk values of any two intelligent vehicles on the candidate prediction trajectory with the highest trajectory prediction probability value; Calculating a sudden high - risk cost value of the target vehicle according to the risk value of each intelligent vehicle and a preset scaling factor on the candidate prediction trajectory with the highest trajectory prediction probability value; Calculating the risk cost value of the target vehicle according to the self - protection risk cost value, the care cost value, and the sudden high - risk cost value of the target vehicle.

7. The trajectory prediction method for an autonomous vehicle according to claim 1, characterized in that, The method further includes: Calculating a loss value between the target trajectory and the candidate prediction trajectory based on a preset loss function, where the loss value includes a smooth loss value of displacement deviation and a cross - entropy loss value of intention recognition; Optimizing a risk optimization module according to the risk cost value, the smooth loss value of displacement deviation, and the cross - entropy loss value of intention recognition to obtain the optimized risk optimization module, where the risk optimization module is used to optimize the candidate prediction trajectory with the highest trajectory prediction probability value.

8. A trajectory prediction device for an autonomous vehicle, characterized in that, Including: An acquisition module, configured to acquire scene map data of a target scene where a target vehicle is located, historical trajectory data of multiple intelligent vehicles in the target scene, and current driving states of the multiple intelligent vehicles, where the intelligent vehicles include the target vehicle and multiple surrounding vehicles; A prediction module, configured to determine multiple candidate prediction trajectories of the target vehicle at a future moment and trajectory prediction probability values corresponding to the multiple candidate prediction trajectories according to the historical trajectory data and the scene map data of the multiple intelligent vehicles; A calculation module, configured to calculate a risk cost value of the target vehicle according to the trajectory prediction probability values corresponding to the multiple candidate prediction trajectories of the target vehicle at the future moment and the current driving states of the multiple intelligent vehicles; An optimization module, configured to optimize the candidate prediction trajectory with the highest trajectory prediction probability value according to the risk cost value of the target vehicle to generate a target prediction trajectory of the target vehicle at the future moment.

9. A terminal device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the method according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and the computer program implements the method according to any one of claims 1 to 7 when executed by a processor.

Images