Vehicle trajectory planning method and device, electronic equipment and storage medium

Abstract

The invention relates to the technical field of automatic driving vehicles, and discloses a vehicle trajectory planning method and device, electronic equipment and a storage medium. The method comprises the steps of collecting at least one kind of environmental perception data; performing space-time alignment and standardization processing on all the environmental perception data to at least fuse data features of all the environmental perception data to generate standard perception data; encoding the standard sensing data according to the data type of the standard sensing data, and performing feature fusion operation on all encoding results to generate global features; selecting a corresponding attention mechanism to process the global features by using a trajectory decoder through a trajectory planning model according to a trajectory generation demand in a current scene so as to generate an initial trajectory plan; the maximum allowable curvature is determined through the real-time vehicle speed and the yaw angle, and a curvature speed coupling loss function is constructed; and optimizing the initial trajectory planning through the curvature velocity coupling loss function to obtain the vehicle trajectory planning. The vehicle use experience of the user is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of autonomous vehicle technology, and in particular to a vehicle trajectory planning method, apparatus, electronic device, and storage medium. Background Technology

[0002] With the rapid development of autonomous driving technology, trajectory planning plays a crucial role in autonomous driving systems. In the field of autonomous driving, imitation learning-based planning methods have shown great potential in handling complex driving scenarios and multimodal driving behaviors. However, current learning-driven planning systems still have limitations when facing complex scenarios, especially in terms of uneven performance across lateral and longitudinal planning tasks. Traditional imitation learning methods are often susceptible to distribution bias (the behavioral strategies learned by the model on training data may not be adaptable to unseen scenarios encountered during testing, leading to unreliable generated trajectories) and causal confusion, resulting in autonomous vehicles generating less flexible and safer planned trajectories when facing complex scenarios.

[0003] Especially in autonomous driving scenarios, the environment is complex and ever-changing, and vehicle trajectories are influenced by multiple factors, making it difficult for imitation learning models to accurately adapt to new scenarios in real-world applications. Causal confusion refers to the fact that imitation learning methods typically extract features directly from environmental observation data without explicitly considering causal relationships, causing the model to potentially ignore important causal information when generating trajectories. For example, the model may not be able to distinguish whether the vehicle's deceleration is due to an obstacle ahead or a problem with its own driving strategy. Furthermore, the planned trajectories of autonomous vehicles must meet physical feasibility requirements, including limits on steering angle, acceleration, and speed. However, trajectories generated by imitation learning often ignore these kinematic and dynamic constraints, potentially resulting in less smooth trajectories in actual driving, thus affecting vehicle safety and comfort. Summary of the Invention

[0004] The purpose of this invention is to provide a vehicle trajectory planning method, device, electronic device, and storage medium to at least solve the problem of unsmooth trajectory switching affecting vehicle safety and comfort during autonomous driving, thereby improving the smoothness of autonomous driving trajectory switching and ensuring the user's driving experience.

[0005] To address the aforementioned technical problems, in a first aspect, the present invention provides a vehicle trajectory planning method, comprising at least:

[0006] Collect at least one type of environmental sensing data;

[0007] All the environmental perception data are spatiotemporally aligned and standardized to fuse the data features of at least all the environmental perception data to generate standard perception data.

[0008] The standard sensing data is encoded according to its data type, and feature fusion is performed on all encoding results to generate global features.

[0009] Based on the trajectory generation requirements in the current scenario, the trajectory planning model utilizes the trajectory decoder to select the corresponding attention mechanism to process the global features, thereby generating at least one initial trajectory plan.

[0010] The maximum permissible curvature is determined by real-time vehicle speed and yaw angle, in order to construct a curvature-velocity coupling loss function;

[0011] The initial trajectory planning is optimized using the curvature-velocity coupling loss function to obtain the vehicle trajectory planning.

[0012] Optionally, the data type includes at least one of static obstacles, vehicle status, vectorized map information, traffic participant trajectories, and global reference paths;

[0013] The encoding results include at least one of the following: static encoding results, vehicle status encoding results, map encoding results, traffic encoding results, and route encoding results.

[0014] Optionally, the step of performing encoding processing on the standard sensing data according to the data type of the standard sensing data, and performing feature fusion operation on all encoding results to generate global features, specifically includes:

[0015] If the data type of the standard sensing data is the static obstacle, then a preset encoder is used to perform encoding processing on the standard sensing data according to the location information and category information of the static obstacle to obtain the static encoding result;

[0016] If the data type of the standard sensing data is the vehicle state, then the standard sensing data is linearly transformed using a preset encoder and encoded using a random masking operation to obtain the vehicle state encoding result.

[0017] If the data type of the standard sensing data is the vectorized map information, then a preset encoder is used to extract point features from the standard sensing data and integrate all the point features to obtain the map encoding result;

[0018] If the data type of the standard perception data is the trajectory of the traffic participant, then the standard perception data is encoded using a preset encoder based on a feature pyramid network of neighborhood attention to obtain the traffic coding result;

[0019] If the data type of the standard sensing data is the global reference path, then the information of each reference path point in the standard sensing data is extracted using a preset encoder, and the path encoding result is generated based on all the reference path point information.

[0020] Time codes are added to the static encoding result, the vehicle status encoding result, the map encoding result, the traffic encoding result, and the route encoding result, respectively. A feature fusion operation is then performed on all encoding results after adding the time codes to generate the global feature.

[0021] Optionally, the step of processing the global features by selecting the corresponding attention mechanism through a trajectory decoder using a trajectory planning model based on the trajectory generation requirements of the current scene to generate at least one initial trajectory plan specifically includes:

[0022] Based on the trajectory generation requirements of the current scenario, the lane line where the current vehicle is located is determined as the lateral reference line;

[0023] A vectorization operation is performed on the horizontal reference line to generate a horizontal query vector based on the vectorization operation result using a preset encoder;

[0024] Establish a vertical learnable embedding matrix, and determine the learnable embedding matrix as the vertical query vector;

[0025] The horizontal query vector and the vertical query vector are concatenated to generate a combined horizontal and vertical query vector;

[0026] The trajectory planning model utilizes the trajectory decoder to select the corresponding attention mechanism based on the comprehensive query vector to process the global features, thereby generating at least one initial trajectory plan.

[0027] Optionally, the step of using the trajectory planning model and the trajectory decoder to select the corresponding attention mechanism based on the comprehensive query vector to process the global features to generate at least one initial trajectory plan specifically includes:

[0028] The trajectory planning model utilizes the trajectory decoder to select a lateral self-attention mechanism based on the comprehensive query vector to obtain the lateral dimension information of the global features;

[0029] The trajectory planning model utilizes the trajectory decoder to select a vertical self-attention mechanism based on the comprehensive query vector to obtain the vertical dimension information of the global features;

[0030] At least one initial trajectory plan is generated based on the horizontal and vertical dimension information using a cross-attention mechanism.

[0031] Optionally, before generating at least one initial trajectory plan by selecting the corresponding attention mechanism through the trajectory decoder using the trajectory planning model according to the trajectory generation requirements of the current scene, the method further includes:

[0032] An initial trajectory planning model, a causal learning loss function, a positive sample generation strategy, and a negative sample generation strategy are constructed, and causal intervention data are collected based on the positive sample generation strategy and the negative sample generation strategy.

[0033] The initial trajectory planning model is trained based on the causal intervention data and the causal learning loss function to obtain the trajectory planning model.

[0034] Secondly, the present invention also provides a vehicle trajectory planning device, comprising at least:

[0035] The information acquisition module is used to collect at least one type of environmental sensing data;

[0036] The standard perception module is used to perform spatiotemporal alignment and standardization processing on all the environmental perception data, so as to fuse the data features of at least all the environmental perception data to generate standard perception data.

[0037] The global feature module is used to perform encoding processing on the standard sensing data according to the data type of the standard sensing data, and to perform feature fusion operation on all encoding results to generate global features;

[0038] The initial trajectory module is used to generate at least one initial trajectory plan by using a trajectory planning model and a trajectory decoder to select the corresponding attention mechanism to process the global features according to the trajectory generation requirements of the current scene.

[0039] The function building module is used to determine the maximum allowable curvature by real-time vehicle speed and yaw angle, so as to construct the curvature-velocity coupling loss function;

[0040] The trajectory determination module is used to optimize the initial trajectory planning through the curvature-velocity coupling loss function to obtain the vehicle trajectory planning.

[0041] Optionally, the data type includes at least one of static obstacles, vehicle status, vectorized map information, traffic participant trajectories, and global reference paths;

[0042] The encoding results include at least one of the following: static encoding results, vehicle status encoding results, map encoding results, traffic encoding results, and route encoding results.

[0043] Thirdly, the present invention also provides an electronic device, including a memory and a processor, the memory storing a computer program executable on the processor, the processor executing the program to implement the steps of the vehicle trajectory planning method according to any one of the first aspects.

[0044] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps in the vehicle trajectory planning method of any one of the first aspects.

[0045] The technical solution provided in this invention firstly collects at least one type of environmental perception data; secondly, it performs spatiotemporal alignment and standardization processing on all environmental perception data to fuse the data features of at least all environmental perception data, generating standard perception data; thirdly, it performs encoding processing on the standard perception data according to the data type of the standard perception data, and performs feature fusion operation on all encoding results to generate global features; fourthly, according to the trajectory generation requirements of the current scenario, it uses a trajectory planning model and a trajectory decoder to select the corresponding attention mechanism to process the global features, generating at least one initial trajectory plan; then, it determines the maximum allowable curvature through real-time vehicle speed and yaw angle to construct a curvature-vehicle coupling loss function; finally, it optimizes the initial trajectory plan through the curvature-vehicle coupling loss function to obtain the vehicle trajectory plan.

[0046] Therefore, this invention, on the one hand, achieves the fusion of all data features by collecting various environmental perception data and performing spatiotemporal alignment and standardization processing. This ensures that the fused standard perception data contains complete environmental information about the vehicle around that point in time, improving the processing efficiency of subsequent encoding. On the other hand, this invention determines the maximum permissible curvature by using real-time vehicle speed and yaw angle to construct a curvature-velocity coupling loss function. This ensures that the generated vehicle trajectory planning strictly conforms to vehicle kinematics and dynamics constraints, guaranteeing the user's driving experience. Attached Figure Description

[0047] Figure 1 This is a flowchart of a vehicle trajectory planning method provided in an embodiment of the present invention;

[0048] Figure 2 This is a sub-flowchart of a vehicle trajectory planning method provided in an embodiment of the present invention;

[0049] Figure 3 This is an architectural diagram of a static obstacle encoder provided in an embodiment of the present invention;

[0050] Figure 4 This is an architecture diagram of a vehicle status encoder provided in an embodiment of the present invention;

[0051] Figure 5 This is an architecture diagram of a map information encoder provided in an embodiment of the present invention;

[0052] Figure 6 This is an architecture diagram of a traffic participant encoder provided in an embodiment of the present invention;

[0053] Figure 7 This is an architecture diagram of a global path encoder provided in an embodiment of the present invention;

[0054] Figure 8 This is a flowchart of another vehicle trajectory planning method provided in an embodiment of the present invention;

[0055] Figure 9 This is an architecture diagram of a trajectory decoder provided in an embodiment of the present invention;

[0056] Figure 10 This is an embedded matrix reference diagram provided in an embodiment of the present invention;

[0057] Figure 11 This is a schematic diagram of the structure of a vehicle trajectory planning device provided in an embodiment of the present invention;

[0058] Figure 12 This is an architectural diagram of a vehicle trajectory planning device provided in an embodiment of the present invention;

[0059] Figure 13 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0061] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to limit the application. The singular forms “a,” “said,” and “the” used in the embodiments of this application and the appended claims are also intended to include the plural forms, and “multiple” generally includes at least two unless the context clearly indicates otherwise.

[0062] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0063] It should be understood that although the terms first, second, third, etc., may be used in the embodiments of this application, these descriptions should not be limited to these terms. These terms are only used to distinguish the descriptions. For example, first may also be referred to as second without departing from the scope of the embodiments of this application, and similarly, second may also be referred to as first.

[0064] Depending on the context, the words “if” or “suppose” as used here can be interpreted as “when” or “in response to determination” or “in response to detection.” Similarly, depending on the context, the phrases “if determination” or “if detection (of the stated condition or event)” can be interpreted as “when determination” or “in response to determination” or “when detection (of the stated condition or event)” or “in response to detection (of the stated condition or event).”

[0065] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or device. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device that includes said element.

[0066] It should be noted that any symbols and / or numbers present in the specification that are not marked in the accompanying drawings are not reference numerals.

[0067] Figure 1 This is a flowchart of a vehicle trajectory planning method provided by an embodiment of the present invention. This embodiment is applicable to trajectory planning scenarios for at least various types of autonomous vehicles. The vehicle trajectory planning method can be, but is not limited to, executed by the vehicle trajectory planning device in this embodiment as the execution subject, which can be implemented in software and / or hardware. Figure 1 As shown, the vehicle trajectory planning method includes at least the following steps:

[0068] S1. Collect at least one type of environmental perception data.

[0069] Environmental perception data can be information about the vehicle's surrounding environment collected using various vehicle sensors (such as LiDAR, onboard cameras, radar, high-precision maps, etc.). The data types and characteristics of data collected by various vehicle sensors are as follows:

[0070] LiDAR (Light Detection and Ranging): This technology generates a 3D point cloud map of the environment by emitting laser beams and receiving reflected signals. Point cloud data can accurately capture the spatial position, distance, and contour information of surrounding objects, making it particularly suitable for detecting static and dynamic obstacles (such as vehicles, pedestrians, and road boundaries). While LiDAR data offers high spatial accuracy, it requires temporal alignment with data from other sensors.

[0071] Vehicle-mounted cameras capture two-dimensional images of the environment, including traffic signs, traffic lights, lane markings, pedestrians, and other vehicles around the vehicle. Camera data has high resolution and rich visual features, but its depth perception capability is relatively limited. Typically, vehicles are equipped with multiple cameras (such as front-view, side-view, and rear-view cameras) to cover the entire surrounding field of vision.

[0072] Radar: Utilizes radio waves to detect the position, speed, and distance of objects, primarily used to obtain relative speed and position information of surrounding vehicles. Radar data is unaffected by lighting conditions and weather conditions, making it suitable for use at night and in adverse weather. While its spatial resolution is relatively low, its speed detection performance is excellent, and it can complement data from lidar and cameras.

[0073] High-precision maps provide static environmental information, including the precise locations of lane lines, road boundaries, traffic signs, and infrastructure. Map data helps vehicles determine their position in a global coordinate system and provides static constraints for trajectory generation. The information in high-precision maps is relatively fixed and unaffected by real-time environmental changes.

[0074] S2. Perform spatiotemporal alignment and standardization on all environmental perception data to fuse at least the data features of all environmental perception data and generate standard perception data.

[0075] Spatiotemporal alignment aims to solve the problem of collaborative processing of temporal and spatial information in multimodal data collected by various vehicle sensors. Its core is to map and match spatiotemporal data from different sources and dimensions through algorithmic models. Standardization can be the transformation of data from different sources, formats, and units into a unified, comparable, and input-to-model format. Standardization can include: (1) Coordinate normalization, transforming all sensor data into the same coordinate system so that all sensors can describe the same location. (2) Time normalization (Time Synchronization + Resampling), due to the different frequencies of each sensor, for example: camera: 30Hz, LiDAR: 10Hz, Radar: 20Hz, interpolation, resampling, etc. are required to ensure that the information of all sensors at the same time t can be used together. (3) Numerical normalization, multimodal data have extremely large scale differences, so numerical normalization is required. (4) Format unification, unifying data of different formats into an intermediate structure that the model can read. Understandably, after performing spatiotemporal alignment and standardization, the data features from cameras, LiDAR, radar, and high-precision maps can be fused into a unified cache, thereby generating standard perception data. This standard perception data contains complete environmental information about the vehicle's surroundings at that point in time, which is used for subsequent scene coding, feature extraction, and other operations.

[0076] S3. Perform encoding processing on the standard sensing data according to the data type of the standard sensing data, and perform feature fusion operation on all encoding results to generate global features.

[0077] The encoding process can be understood as using an encoder to extract features from standard perceived data. In one specific implementation, optionally, the data type includes at least one of static obstacles, vehicle state, vectorized map information, traffic participant trajectories, and global reference paths; the encoding result includes at least one of static encoding result, vehicle state encoding result, map encoding result, traffic encoding result, and path encoding result. It is understood that the encoder can use different encoding methods to encode various input features such as static obstacles, vehicle state, vectorized map information, traffic participant trajectory features, and global reference paths, thus fully capturing the unique information of each type of data feature.

[0078] S4. Based on the trajectory generation requirements in the current scenario, the trajectory planning model uses the trajectory decoder to select the corresponding attention mechanism to process global features in order to generate at least one initial trajectory plan.

[0079] The trajectory generation requirement could be that the autonomous vehicle needs to change lanes or overtake. The trajectory planning model could be a pre-trained neural network model.

[0080] S5. Determine the maximum permissible curvature by using real-time vehicle speed and yaw angle to construct a curvature-velocity coupling loss function.

[0081] The yaw angle is the angle between the vehicle's direction of travel and a reference coordinate system (such as the road coordinate system). It reflects the vehicle's attitude and is directly related to steering behavior. The vehicle's lateral acceleration can be accurately calculated using the real-time vehicle speed and the yaw rate.

[0082] S6. Optimize the initial trajectory planning through the curvature-velocity coupling loss function to obtain the vehicle trajectory planning.

[0083] In this context, optimization can be understood as constraining the turning limits of the initial trajectory planning to ensure that the vehicle trajectory planning is within the kinematic and dynamic constraints of the actual vehicle. Therefore, this embodiment introduces a differentiable curvature-velocity coupling constraint layer in the output stage of vehicle trajectory planning. First, a lateral acceleration 'a' is established based on the bicycle model. y The relationship between a and the trajectory curvature y ≈v 2 κ, and set the maximum permissible lateral acceleration a based on tire-road adhesion conditions. y,max =ημg2. Where μ represents the tire-road adhesion coefficient, η is the safety margin coefficient (obtained through calibration), and g represents gravitational acceleration. From this, the upper bound of the maximum theoretical curvature varying with vehicle speed can be obtained. :

[0084] ;

[0085] In the formula, This is a numerical stability term. Furthermore, to account for the influence of vehicle attitude on stability, a heading angle is introduced. tangential direction of the road deviation And perform attitude-dependent contraction on the upper bound of maximum curvature:

[0086] ;

[0087] In the formula, This represents the maximum permissible heading deviation. These are the weighting coefficients. This indicates ReLU truncation.

[0088] Subsequently, the discrete trajectory P generated by the decoder t = (x t ,y tInstantaneous curvature is calculated using differentiable geometric relationships:

[0089] ;

[0090] When the actual curvature of the predicted trajectory exceeds the maximum permissible curvature, it is penalized by a curvature constraint loss function, thereby automatically reducing the degree of local curvature through backpropagation and ensuring that the generated trajectory satisfies the vehicle dynamics feasible region. The final curvature-velocity coupling loss function L... k The expression can be:

[0091] ;

[0092] In the formula, k represents curvature, K max κ represents the maximum permissible curvature, T represents the time step range, t represents the time, and κ represents the maximum permissible curvature. t The trajectory curvature at time t. Through the aforementioned differentiable constraints, the decoder can explicitly follow vehicle dynamics limitations during training, ensuring the generated trajectory remains controllable and physically feasible even under extreme conditions such as high speeds and high yaw angles. In the trajectory post-processing stage, a real-time quadratic programming (QP) optimizer (an optimization algorithm) can be used to further optimize the initial trajectory planning, using the initial trajectory as a reference and acceleration... and urgency As a hard constraint, the trajectory point coordinates and velocity profile are iteratively adjusted to ensure trajectory smoothness and dynamic feasibility. It is known that 'a' represents acceleration, and 'a'... t a represents the acceleration at time t. max Indicates the maximum acceleration limit, jerk t It represents the jerk (rate of change of acceleration) at time t.

[0093] The technical solution provided in this embodiment firstly collects at least one type of environmental perception data; secondly, it performs spatiotemporal alignment and standardization processing on all environmental perception data to fuse the data features of at least all environmental perception data, generating standard perception data; thirdly, it performs encoding processing on the standard perception data according to the data type of the standard perception data, and performs feature fusion operation on all encoding results to generate global features; fourthly, according to the trajectory generation requirements in the current scenario, it uses a trajectory planning model and a trajectory decoder to select the corresponding attention mechanism to process the global features, generating at least one initial trajectory plan; then, it determines the maximum allowable curvature through real-time vehicle speed and yaw angle to construct a curvature-vehicle coupling loss function; finally, it optimizes the initial trajectory plan through the curvature-vehicle coupling loss function to obtain the vehicle trajectory plan.

[0094] Therefore, this embodiment achieves the fusion of all data features by collecting various environmental perception data and performing spatiotemporal alignment and standardization. This ensures that the fused standard perception data contains complete environmental information about the vehicle at that point in time, improving the processing efficiency of subsequent encoding. Furthermore, this embodiment determines the maximum permissible curvature by using real-time vehicle speed and yaw angle to construct a curvature-velocity coupling loss function. This ensures that the generated vehicle trajectory planning strictly conforms to vehicle kinematics and dynamics constraints, guaranteeing a superior user experience.

[0095] Based on the above embodiments or implementation methods Figure 2 This is a sub-flowchart of a vehicle trajectory planning method provided in an embodiment of the present invention. This embodiment refines step S3 based on the above embodiment. Figure 2 As shown, step S3 includes at least the following steps:

[0096] S31. If the data type of the standard sensing data is static obstacles, then the standard sensing data is encoded using a preset encoder based on the location and category information of the static obstacles to obtain static encoding results.

[0097] Static obstacles can include buildings, traffic signs, etc. Figure 3 This is an architectural diagram of a static obstacle encoder provided in an embodiment of the present invention, as shown below. Figure 3 As shown, the preset encoder in step S31 can be a multilayer perceptron (MLP). The MLP includes the following two layers: a Fourier embedding layer: The Fourier embedding layer takes the input static obstacle's position information (coordinates in the vehicle's coordinate system) as input and transforms it into a high-dimensional representation so that the subsequent model can better identify the spatial location of the obstacle. A category embedding layer: The category embedding layer takes the input static obstacle's category information (e.g., buildings, trees, road boundaries) as input and transforms the category information into an embedding representation, thereby obtaining the static encoding result. This layer helps the trajectory planning model distinguish different types of obstacles, thus generating the corresponding trajectory planning strategy more accurately. It is known that the encoding results can all be represented in the form of embedding vectors.

[0098] S32. If the data type of the standard sensing data is the vehicle state, then the standard sensing data is linearly transformed using a preset encoder and encoded using a random masking operation to obtain the vehicle state encoding result.

[0099] The vehicle status can include information such as vehicle speed and acceleration. Figure 4 This is an architecture diagram of a vehicle status encoder provided in an embodiment of the present invention, such as... Figure 4As shown, it can be understood that for encoding the vehicle state, this module first performs a linear transformation on the input data to standardize the data scale of the standard perceived data of the vehicle state. Then, it performs position embedding and random masking operations, and finally obtains the relationship between inputs at different positions through a self-attention mechanism. Specifically, the vehicle state (such as speed, acceleration, and direction) is first input into a linear layer and linearly transformed. This process helps standardize the vehicle state information, making it suitable for subsequent embedding operations. After passing through the position embedding layer, the vehicle state can better retain its positional relationship within the overall vehicle state. To prevent the trajectory planning model from overfitting to the vehicle state information, a random masking operation is added. This operation randomly masks some input features, prompting the model to focus on more global information and improving generalization ability. Finally, the self-attention mechanism captures the relationship between the vehicle state at different positions, enabling the encoder to effectively identify the interaction information of various parts of the vehicle state and improve the expressive power of the encoding.

[0100] S33. If the data type of the standard sensing data is vectorized map information, then the point features in the standard sensing data are extracted using a preset encoder, and all point features are integrated to obtain the map encoding result.

[0101] in, Figure 5 This is an architecture diagram of a map information encoder provided in an embodiment of the present invention, such as... Figure 5 As shown, the preset encoder in step S33 can be a PointNet-based encoding method. This encoder can effectively process point features in map information and integrate the features of multiple points into a global feature representation. It can be understood that by extracting the features of each point in the vectorized map and utilizing the local operations of PointNet, the encoder can capture local information such as road boundaries and lane lines in the map. Based on the point feature extraction, the max pooling layer of the preset encoder can globally integrate the features of all points to generate a global feature vector. It is evident that the global feature vector contains the overall features of the vectorized map information, providing static constraint information for trajectory planning.

[0102] S34. If the data type of the standard perception data is the trajectory of traffic participants, then the standard perception data is encoded using a preset encoder based on a feature pyramid network of neighborhood attention to obtain traffic coding results.

[0103] Among them, the trajectory of traffic participants can be the historical trajectory of traffic participants. Figure 6 This is an architecture diagram of a traffic participant encoder provided in an embodiment of the present invention, such as... Figure 6As shown, it is understandable that, in order to extract useful information from the historical trajectories of traffic participants, the pre-encoder can use a feature pyramid network based on neighborhood attention to encode the trajectory data. This network can aggregate historical trajectory information at different scales to generate an embedded representation for each traffic participant. Furthermore, this network can extract trajectory features at different scales through hierarchical aggregation. Lower-level features focus on smaller local details, while higher-level features focus on a wider range of trajectory information. Furthermore, to obtain local interaction information while reducing computational cost, the pre-encoder can further introduce a neighborhood attention mechanism. This mechanism reduces the computational complexity of global attention through local weighting, while effectively capturing the interaction relationships between traffic participants, thus providing more accurate input for trajectory planning.

[0104] S35. If the data type of the standard sensing data is a global reference path, then the information of each reference path point in the standard sensing data is extracted using a preset encoder, and a path coding result is generated based on all the reference path point information.

[0105] The global reference path can be the location information of vehicles on the autonomous driving planning path. Figure 7 This is an architecture diagram of a global path encoder provided in an embodiment of the present invention, such as... Figure 7 As shown, the global reference path encoder uses an MLP for feature extraction (i.e., encoding). This encoding method helps the trajectory planning model capture the vehicle's position information on the planned path. By inputting the points of the reference path into the multilayer perceptron, global path information can be extracted. Step S35 helps generate the vehicle's position vector in the path, providing target direction constraints for trajectory generation.

[0106] S36. Add time codes to the static coding results, vehicle status coding results, map coding results, traffic coding results and route coding results respectively, and perform feature fusion operation on all coding results after adding time codes to generate global features.

[0107] Understandably, to capture the spatiotemporal relationships of different encoding results, the pre-defined encoder adds temporal encoding to each embedding vector (i.e., each encoding result) to ensure the model can distinguish the temporal and spatial features of each input. By using scene temporal encoding, all encoding vectors can be integrated into the final encoder's output feature vector (i.e., global features). Scene temporal encoding embeds the spatiotemporal information of each input feature into the final feature representation. Specifically, temporal encoding allows the trajectory planning model to identify the correlation between data at different points in time, while spatial features help the trajectory planning model obtain the spatial location of each feature, thereby generating a global spatiotemporal embedding representation. The embedding vectors processed by scene temporal encoding are fused to generate a high-dimensional global feature representation. This feature includes integrated information on static and dynamic obstacles, vehicle state, reference path, and traffic participant trajectories, providing comprehensive spatiotemporal environmental awareness for trajectory generation.

[0108] Based on the above embodiments or implementation methods Figure 8 This is a flowchart of another vehicle trajectory planning method provided by an embodiment of the present invention. This embodiment is based on the above embodiment with additions and refinements. Figure 8 As shown, the vehicle trajectory planning method includes at least the following steps:

[0109] S1. Collect at least one type of environmental perception data.

[0110] S2. Perform spatiotemporal alignment and standardization on all environmental perception data to fuse at least the data features of all environmental perception data and generate standard perception data.

[0111] S3. Perform encoding processing on the standard sensing data according to the data type of the standard sensing data, and perform feature fusion operation on all encoding results to generate global features.

[0112] S71. Construct an initial trajectory planning model, a causal learning loss function, a positive sample generation strategy, and a negative sample generation strategy, and collect causal intervention data based on the positive sample generation strategy and the negative sample generation strategy.

[0113] The causal learning loss function is used to enhance the trajectory planning model's understanding of driving causal relationships and suppress trajectory generation risks caused by distributional bias, thereby improving the causal inference ability and robustness of the trajectory planning model. Causal intervention data can be datasets that conform to either positive or negative sample generation strategies.

[0114] Positive sample generation strategies mainly include state perturbation and non-interactive agent removal strategies. State perturbation involves injecting speed fluctuations that conform to vehicle dynamics (such as within ±10% of the boundary) into the expert trajectory to preserve causal events (such as steering signals before lane changes). Non-interactive agent removal involves randomly deleting obstacles that have no interaction with the main vehicle (such as vehicles in the opposite lane) to enhance the model's attention to key interactive targets.

[0115] Negative sample generation strategies mainly involve causal disruption operations, such as forward agent removal: deleting key obstacles (such as cutting-in vehicles) to create the illusion of "no collision risk"; interactive agent insertion: adding virtual obstacles outside the safe distance to simulate accidental braking scenarios, etc.

[0116] S72. Train an initial trajectory planning model based on causal intervention data and a causal learning loss function to obtain the trajectory planning model.

[0117] In this process, after acquiring causal intervention data, in addition to the standard autonomous driving dataset, a causal learning loss function is added during the initial trajectory planning model training to train the initial trajectory planning model's understanding of driving causal relationships. The expression for the causal learning loss function can be:

[0118] ;

[0119] Among them, L enhance This represents the causal learning loss function. N represents the sample batch size. The scene embedding vector representing the positive sample (scene semantics after intervention by the positive sample generation strategy) represents the scene semantic representation obtained after the intervention of the positive sample generation strategy, which conforms to the causal logic of traffic and is used to characterize the causal relationship behind real driving behavior. E is the original positive sample scene embedding vector (i.e., without intervention), representing the scene features extracted from the original normal driving scene. The model aims to approximate its semantics. . The scene embedding vector representing a negative sample (not conforming to causal logic) is the scene representation obtained after intervening with the negative sample generation strategy, which does not conform to traffic behavior patterns and is used as a causal opposition sample. T represents the original scene embedding vector of the negative sample, the original scene features of the negative sample, used to compare with... Combine and construct a counter-causal comparison term. This represents the cosine similarity of positive samples; the higher the similarity, the closer the two scenarios are. It is a temperature coefficient used to control the sharpness of the similarity distribution, mainly used to adjust the distinguishing strength between positive and negative samples in contrastive learning.

[0120] S41. Based on the trajectory generation requirements of the current scenario, the lane line where the current vehicle is located is determined as the lateral reference line.

[0121] in, Figure 9 This is an architecture diagram of a trajectory decoder provided in an embodiment of the present invention. See also... Figure 9 To alleviate the performance imbalance between lateral and longitudinal planning tasks, this embodiment employs a query-based decoding method to support the lateral and longitudinal planning requirements of autonomous driving. Steps S41-S45 can be understood as the decoding process of global features using a trajectory decoder. The trajectory decoder first constructs three queries (lateral query, longitudinal query, and combined lateral-longitudinal query) for different trajectory generation needs. The lateral query is mainly used to obtain the vehicle's lateral information to support the vehicle's lateral driving behavior within the lane, including following or changing lanes. The longitudinal query is used to capture the vehicle's longitudinal dynamic information, especially the vehicle's acceleration and deceleration behavior over time. The combined query performs a simple linear fusion of the feature vectors from the lateral and longitudinal queries to generate a new query vector. The combined query focuses on the comprehensive impact of the entire scene on the vehicle's trajectory, ensuring that trajectory generation can be effectively balanced in both lateral and longitudinal dimensions to meet overall driving requirements.

[0122] S42. Perform a vectorization operation on the horizontal reference line to generate a horizontal query vector based on the result of the vectorization operation using a preset encoder.

[0123] The purpose of vectorization is to facilitate the encoding of the horizontal reference lines. The preset encoder in step S42 can be a PointNet-like encoder. The horizontal query vector can be generated by the PointNet-like encoder.

[0124] S43. Establish a vertical learnable embedding matrix and determine the learnable embedding matrix as the vertical query vector.

[0125] in, Figure 10 This is an embedded matrix reference diagram provided in an embodiment of the present invention, such as... Figure 10 As shown, regions of equal size are resampled vertically, and an equal number of learnable embedding matrices are built based on the number of generated regions, serving as the vertical query vector.

[0126] S44. Concatenate the horizontal query vector and the vertical query vector to generate a combined horizontal and vertical query vector.

[0127] S45. Using a trajectory planning model, the trajectory decoder selects the corresponding attention mechanism based on the comprehensive query vector to process global features, so as to generate at least one initial trajectory plan.

[0128] The trajectory decoder consists of multiple decoding layers (L layers in total). Each decoding layer processes input features through three attention mechanisms (lateral self-attention, vertical self-attention, and cross-attention based on the queried scene) and progressively generates trajectory predictions. In one specific implementation, step S45 may optionally include:

[0129] (45-1) By using the trajectory planning model and the trajectory decoder to select the lateral self-attention mechanism based on the comprehensive query vector, the lateral dimension information of the global features is obtained.

[0130] The role of the lateral self-attention mechanism is to capture detailed information about lane keeping and lane changing behaviors and provide lateral dimension information for trajectory generation.

[0131] (45-2) The trajectory planning model utilizes the trajectory decoder to select the vertical self-attention mechanism based on the comprehensive query vector to obtain the vertical dimension information of the global features.

[0132] The longitudinal query vector acquires longitudinal dimension interactions through a longitudinal self-attention mechanism. This mechanism helps the decoder understand how the vehicle accelerates or decelerates over time, especially when there are obstacles or other vehicles ahead, adjusting speed to ensure a safe following distance. It is understood that after step (45-2) in this embodiment, an adaptive query vector can also be output. This adaptive query vector is a high-dimensional vector representation, not a fixed parameter, but a predicted state representation that is adaptively updated according to the scene. As the self-attention mechanism continuously processes lateral / longitudinal / environmental information, the adaptive query vector is dynamically corrected. After being updated by lateral and longitudinal self-attention mechanisms, the adaptive query vector essentially possesses an understanding of the current traffic environment and its own movement trend, representing a dynamically updated internal representation of trajectory generation. Furthermore, as lateral interaction information updates, lateral self-attention allows the adaptive query vector to understand information relevant to lateral movement, such as current lane geometry constraints, the influence of adjacent vehicles, whether lane changing is necessary, and lateral safety distance. Therefore, the lateral query can obtain lateral semantics. Furthermore, as the longitudinal interaction information is updated, the longitudinal self-attention allows the adaptive query vector to understand the information needed for longitudinal motion, such as the distance to the vehicle ahead, acceleration / deceleration trends, longitudinal collision risk, whether to slow down, and acceleration. Therefore, the longitudinal query can obtain longitudinal semantics. Thus, the adaptive query vector can be understood as the latent feature after both lateral and longitudinal updates.

[0133] Further, see also Figure 9Furthermore, the integrated query vector can interact with the multimodal spatiotemporal coding results in real time, enabling deep interaction between the adaptive query vector and the multimodal spatiotemporal coding results. This process allows the adaptive query vector to proactively extract the most relevant environmental information from the scene for current trajectory generation, thereby outputting an adaptive embedding vector that integrates multimodal environmental semantics. This embedding vector not only includes the vehicle's own motion trend but also encodes safety constraints from the surrounding environment, thus effectively guiding the decoder to generate safer and more reasonable future trajectories that avoid collisions, comply with road constraints, and are more reasonable in complex scenes.

[0134] (45-3) Generate at least one initial trajectory plan based on horizontal and vertical dimension information through cross-attention mechanism.

[0135] The cross-attention mechanism integrates the query vector with multimodal spatiotemporal encoding results (i.e., global features), outputting an adaptive embedding vector that incorporates multimodal spatiotemporal feature fusion. This mechanism enables the decoder to identify potential hazards in the environment and generate safer trajectories. By combining vehicle trajectories with environmental features, cross-attention helps the trajectory planning model generate collision-avoiding safe trajectories in complex scenarios, ensuring that the vehicle's path conforms to actual road conditions. Finally, a softmax layer outputs different initial trajectory plans and their corresponding probabilities.

[0136] S5. Determine the maximum permissible curvature by using real-time vehicle speed and yaw angle to construct a curvature-velocity coupling loss function.

[0137] S6. Optimize the initial trajectory planning through the curvature-velocity coupling loss function to obtain the vehicle trajectory planning.

[0138] The technical solution provided in this embodiment firstly collects at least one type of environmental perception data. Further, all environmental perception data undergoes spatiotemporal alignment and standardization processing to fuse the data features of at least all environmental perception data, generating standard perception data. Further, the standard perception data is encoded according to its data type, and feature fusion is performed on all encoding results to generate global features. Further, an initial trajectory planning model, a causal learning loss function, a positive sample generation strategy, and a negative sample generation strategy are constructed, and causal intervention data is collected based on these strategies. Further, the initial trajectory planning model is trained based on the causal intervention data and the causal learning loss function to obtain the trajectory planning model. Further, according to the trajectory generation requirements of the current scenario, the lane line where the current vehicle is located is determined as the lateral reference line. Further, a vectorization operation is performed on the lateral reference line to generate a lateral query vector using a preset encoder based on the vectorization operation results. Further, a vertical learnable embedding matrix is ​​established, and this matrix is ​​determined as the vertical query vector. Further, the lateral query vector and the vertical query vector are concatenated to generate a combined lateral and vertical query vector. Furthermore, the trajectory planning model utilizes a trajectory decoder to select the corresponding attention mechanism based on the comprehensive query vector to process global features, thereby generating at least one initial trajectory plan. Further, the maximum permissible curvature is determined using real-time vehicle speed and yaw angle to construct a curvature-vehicle coupling loss function. Finally, the initial trajectory plan is optimized using the curvature-vehicle coupling loss function to obtain the final vehicle trajectory plan.

[0139] Therefore, this embodiment, on the one hand, aims to fully extract multimodal spatiotemporal feature data. The input sources for this embodiment are target detection results based on Bird's Eye View (BEV), positioning information from the Global Navigation Satellite System (GNSS), vehicle attitude information from the Inertial Measurement Unit (IMU), and high-precision maps. Different encoding methods are designed for spatiotemporal data with different features. Furthermore, to achieve the fusion of different spatiotemporal data in the same scene, an attention mechanism is used to interact with the encoded results, generating a multimodal spatiotemporal feature embedding tensor for subsequent decoding. On the other hand, to correctly guide the trajectory planned by the autonomous driving system, this embodiment designs a trajectory decoder based on query vectors. This decoder, based on a query vector generated by an adaptive query mechanism, interacts with the multimodal spatiotemporal feature embedding tensor output by the encoder through a cross-attention mechanism, ultimately generating the planned trajectory. The generated query vector can simultaneously focus on both lateral and longitudinal trajectory information. On another front, to improve the causal understanding and scene adaptability of the autonomous driving system in complex scenarios, this embodiment designs a data augmentation method for causal intervention. By modifying some data and removing key factors, the training of the model is enhanced. Through this module, the system can generate expected trajectories in different driving situations, effectively reducing the risks caused by distribution deviations.

[0140] Figure 11 This is a schematic diagram of a vehicle trajectory planning device provided in an embodiment of the present invention. This embodiment is applicable to trajectory planning scenarios for at least various types of autonomous vehicles. Figure 12 This is an architectural diagram of a vehicle trajectory planning device provided in an embodiment of the present invention. This vehicle trajectory planning device can be implemented using software and / or hardware. Figure 11 and Figure 12 As shown, the vehicle trajectory planning device includes at least:

[0141] Information acquisition module 110 is used to acquire at least one type of environmental perception data.

[0142] The standard perception module 120 is used to perform spatiotemporal alignment and standardization processing on all environmental perception data, so as to fuse the data features of at least all environmental perception data to generate standard perception data.

[0143] The global feature module 130 is used to perform encoding processing on the standard sensing data according to the data type of the standard sensing data, and to perform feature fusion operation on all encoding results to generate global features.

[0144] The initial trajectory module 140 is used to generate at least one initial trajectory plan by using a trajectory decoder to process global features through a trajectory planning model and selecting the corresponding attention mechanism.

[0145] Function construction module 150 is used to determine the maximum permissible curvature by real-time vehicle speed and yaw angle, so as to construct a curvature-velocity coupling loss function.

[0146] The trajectory determination module 160 is used to optimize the initial trajectory planning through the curvature-velocity coupling loss function to obtain the vehicle trajectory planning.

[0147] Optionally, the data type includes at least one of the following: static obstacles, vehicle status, vectorized map information, traffic participant trajectories, and global reference paths;

[0148] The encoding results include at least one of the following: static encoding results, vehicle status encoding results, map encoding results, traffic encoding results, and route encoding results.

[0149] Optionally, the global feature module 130 is specifically used for:

[0150] When the standard perception data is of static obstacles, a preset encoder is used to encode the standard perception data based on the location and category information of the static obstacles to obtain static encoding results. When the standard perception data is of vehicle status, a preset encoder is used to perform a linear transformation on the standard perception data and complete the encoding process through random masking to obtain vehicle status encoding results. When the standard perception data is of vectorized map information, a preset encoder is used to extract point features from the standard perception data and integrate all point features to obtain map encoding results. When the standard perception data is of traffic participant trajectories, a preset encoder is used to perform encoding operations on the standard perception data based on a feature pyramid network with neighborhood attention to obtain traffic encoding results. When the standard perception data is of global reference paths, a preset encoder is used to extract information from each reference path point in the standard perception data and generate path encoding results based on all reference path point information. Furthermore, time-based encoding is added to the static encoding results, vehicle status encoding results, map encoding results, traffic encoding results, and path encoding results, and feature fusion is performed on all encoding results after adding time-based encoding to generate global features.

[0151] Optionally, the initial trajectory module 140 is specifically used for:

[0152] Based on the trajectory generation requirements of the current scenario, the lane line where the current vehicle is located is determined as the lateral reference line; and a vectorization operation is performed on the lateral reference line to generate a lateral query vector using a preset encoder based on the vectorization operation result; a vertical learnable embedding matrix is ​​established and determined as the vertical query vector; the lateral query vector and the vertical query vector are concatenated to generate a combined lateral and vertical query vector; and a trajectory planning model is used to select the corresponding attention mechanism based on the combined query vector to process global features through a trajectory decoder to generate at least one initial trajectory plan.

[0153] Optionally, the initial trajectory module 140 is also specifically used for:

[0154] The trajectory planning model utilizes a trajectory decoder to select a lateral self-attention mechanism based on the comprehensive query vector to obtain the lateral dimension information of the global features; and the trajectory planning model utilizes a trajectory decoder to select a vertical self-attention mechanism based on the comprehensive query vector to obtain the vertical dimension information of the global features; and at least one initial trajectory plan is generated based on the lateral and vertical dimension information through a cross-attention mechanism.

[0155] Optionally, it also includes:

[0156] The model training module 170 is used to construct an initial trajectory planning model, a causal learning loss function, a positive sample generation strategy, and a negative sample generation strategy, and to collect causal intervention data based on the positive sample generation strategy and the negative sample generation strategy; and to train the initial trajectory planning model based on the causal intervention data and the causal learning loss function to obtain the trajectory planning model.

[0157] The technical solution provided in this embodiment firstly acquires at least one type of environmental perception data through an information acquisition module. Further, a standard perception module performs spatiotemporal alignment and standardization processing on all environmental perception data to fuse the data features of at least all environmental perception data, generating standard perception data. Further, a global feature module encodes the standard perception data according to its data type and performs feature fusion on all encoded results to generate global features. Further, an initial trajectory module processes the global features using a trajectory planning model and a trajectory decoder, based on the trajectory generation requirements of the current scene, to generate at least one initial trajectory plan. Further, a function construction module determines the maximum permissible curvature based on real-time vehicle speed and yaw angle to construct a curvature-vehicle coupling loss function. Finally, a trajectory determination module optimizes the initial trajectory plan based on the curvature-vehicle coupling loss function to obtain the vehicle trajectory plan.

[0158] Therefore, this embodiment achieves the fusion of all data features by collecting various environmental perception data and performing spatiotemporal alignment and standardization. This ensures that the fused standard perception data contains complete environmental information about the vehicle at that point in time, improving the processing efficiency of subsequent encoding. Furthermore, this embodiment determines the maximum permissible curvature by using real-time vehicle speed and yaw angle to construct a curvature-velocity coupling loss function. This ensures that the generated vehicle trajectory planning strictly conforms to vehicle kinematics and dynamics constraints, guaranteeing a superior user experience.

[0159] This embodiment provides an electronic device. Figure 13 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. See also: Figure 13 The electronic device 1000 includes a processor 1001 and a memory 1002. The memory 1002 stores computer-readable instructions. When the computer-readable instructions are executed by the processor 1001, the steps in any of the vehicle trajectory planning methods described above are performed. Through the above technical solution, the processor 1001 and the memory 1002 are interconnected and communicate with each other through a communication bus and / or other forms of connection mechanism (not shown). The memory 1002 stores a computer program executable by the processor. When the electronic device 1000 is running, the processor 1001 executes the computer program to execute the vehicle trajectory planning method in any optional implementation of the above embodiments, so as to achieve at least the following functions: collecting at least one type of environmental perception data; performing spatiotemporal alignment and standardization processing on all environmental perception data to fuse at least the data features of all environmental perception data to generate standard perception data; performing encoding processing on the standard perception data according to the data type of the standard perception data, and performing feature fusion operation on all encoding results to generate global features; according to the trajectory generation requirements in the current scenario, using the trajectory planning model and the trajectory decoder to select the corresponding attention mechanism to process the global features to generate at least one initial trajectory plan; determining the maximum allowable curvature through real-time vehicle speed and yaw angle to construct a curvature-vehicle coupling loss function; optimizing the initial trajectory plan through the curvature-vehicle coupling loss function to obtain the vehicle trajectory plan.

[0160] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the vehicle trajectory planning method provided in all embodiments of this application: collecting at least one type of environmental perception data; performing spatiotemporal alignment and standardization on all environmental perception data to fuse the data features of at least all environmental perception data to generate standard perception data; performing encoding processing on the standard perception data according to its data type, and performing feature fusion operation on all encoding results to generate global features; based on the trajectory generation requirements of the current scenario, using a trajectory planning model and a trajectory decoder to select the corresponding attention mechanism to process the global features to generate at least one initial trajectory plan; determining the maximum allowable curvature through real-time vehicle speed and yaw angle to construct a curvature-vehicle coupling loss function; and optimizing the initial trajectory plan through the curvature-vehicle coupling loss function to obtain the vehicle trajectory plan.

[0161] Any combination of one or more computer-readable media may be used. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device.

[0162] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including—but not limited to—electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of transmitting, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0163] The program code contained on a computer-readable medium may be transmitted using any suitable medium, including—but not limited to—wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0164] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—as well as conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0165] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A vehicle trajectory planning method, characterized in that, At least including: Collect at least one type of environmental sensing data; All the environmental perception data are spatiotemporally aligned and standardized to fuse the data features of at least all the environmental perception data to generate standard perception data. The standard sensing data is encoded according to its data type, and feature fusion is performed on all encoding results to generate global features. Based on the trajectory generation requirements in the current scenario, the trajectory planning model utilizes the trajectory decoder to select the corresponding attention mechanism to process the global features, thereby generating at least one initial trajectory plan. The maximum permissible curvature is determined by real-time vehicle speed and yaw angle, in order to construct a curvature-velocity coupling loss function; The initial trajectory planning is optimized using the curvature-velocity coupling loss function to obtain the vehicle trajectory planning.

2. The vehicle trajectory planning method according to claim 1, characterized in that, The data type includes at least one of the following: static obstacles, vehicle status, vectorized map information, traffic participant trajectories, and global reference paths; The encoding results include at least one of the following: static encoding results, vehicle status encoding results, map encoding results, traffic encoding results, and route encoding results.

3. The vehicle trajectory planning method according to claim 2, characterized in that, The step of encoding the standard sensing data according to its data type and performing feature fusion on all encoding results to generate global features specifically includes: If the data type of the standard sensing data is the static obstacle, then a preset encoder is used to perform encoding processing on the standard sensing data according to the location information and category information of the static obstacle to obtain the static encoding result; If the data type of the standard sensing data is the vehicle state, then the standard sensing data is linearly transformed using a preset encoder and encoded using a random masking operation to obtain the vehicle state encoding result. If the data type of the standard sensing data is the vectorized map information, then a preset encoder is used to extract point features from the standard sensing data and integrate all the point features to obtain the map encoding result; If the data type of the standard perception data is the trajectory of the traffic participant, then the standard perception data is encoded using a preset encoder based on a feature pyramid network of neighborhood attention to obtain the traffic coding result; If the data type of the standard sensing data is the global reference path, then the information of each reference path point in the standard sensing data is extracted using a preset encoder, and the path encoding result is generated based on all the reference path point information. Time codes are added to the static encoding result, the vehicle status encoding result, the map encoding result, the traffic encoding result, and the route encoding result, respectively. A feature fusion operation is then performed on all encoding results after adding the time codes to generate the global feature.

4. The vehicle trajectory planning method according to claim 2, characterized in that, The step of generating at least one initial trajectory plan based on the trajectory generation requirements of the current scenario, specifically includes: using a trajectory planning model and a trajectory decoder to select the corresponding attention mechanism to process the global features, thereby generating the initial trajectory plan. Based on the trajectory generation requirements of the current scenario, the lane line where the current vehicle is located is determined as the lateral reference line; A vectorization operation is performed on the horizontal reference line to generate a horizontal query vector based on the vectorization operation result using a preset encoder; Establish a vertical learnable embedding matrix, and determine the learnable embedding matrix as the vertical query vector; The horizontal query vector and the vertical query vector are concatenated to generate a combined horizontal and vertical query vector; The trajectory planning model utilizes the trajectory decoder to select the corresponding attention mechanism based on the comprehensive query vector to process the global features, thereby generating at least one initial trajectory plan.

5. The vehicle trajectory planning method according to claim 4, characterized in that, The step of using the trajectory planning model and the trajectory decoder to select the corresponding attention mechanism based on the comprehensive query vector to process the global features in order to generate at least one initial trajectory plan specifically includes: The trajectory planning model utilizes the trajectory decoder to select a lateral self-attention mechanism based on the comprehensive query vector to obtain the lateral dimension information of the global features; The trajectory planning model utilizes the trajectory decoder to select a vertical self-attention mechanism based on the comprehensive query vector to obtain the vertical dimension information of the global features; At least one initial trajectory plan is generated based on the horizontal and vertical dimension information using a cross-attention mechanism.

6. The vehicle trajectory planning method according to claim 1, characterized in that, Before generating at least one initial trajectory plan by using a trajectory planning model and a trajectory decoder to select the corresponding attention mechanism to process the global features according to the trajectory generation requirements of the current scene, the method further includes: An initial trajectory planning model, a causal learning loss function, a positive sample generation strategy, and a negative sample generation strategy are constructed, and causal intervention data are collected based on the positive sample generation strategy and the negative sample generation strategy. The initial trajectory planning model is trained based on the causal intervention data and the causal learning loss function to obtain the trajectory planning model.

7. A vehicle trajectory planning device, characterized in that, At least including: The information acquisition module is used to collect at least one type of environmental sensing data; The standard perception module is used to perform spatiotemporal alignment and standardization processing on all the environmental perception data, so as to fuse the data features of at least all the environmental perception data to generate standard perception data. The global feature module is used to perform encoding processing on the standard sensing data according to the data type of the standard sensing data, and to perform feature fusion operation on all encoding results to generate global features; The initial trajectory module is used to generate at least one initial trajectory plan by using a trajectory planning model and a trajectory decoder to select the corresponding attention mechanism to process the global features according to the trajectory generation requirements of the current scene. The function building module is used to determine the maximum allowable curvature by real-time vehicle speed and yaw angle, so as to construct the curvature-velocity coupling loss function; The trajectory determination module is used to optimize the initial trajectory planning through the curvature-velocity coupling loss function to obtain the vehicle trajectory planning.

8. The vehicle trajectory planning device according to claim 7, characterized in that, The data type includes at least one of the following: static obstacles, vehicle status, vectorized map information, traffic participant trajectories, and global reference paths; The encoding results include at least one of the following: static encoding results, vehicle status encoding results, map encoding results, traffic encoding results, and route encoding results.

9. An electronic device comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the vehicle trajectory planning method according to any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps in the vehicle trajectory planning method according to any one of claims 1 to 6.

Images