Trajectory planning method based on multi-stage optimization strategy and mixed diffusion model

Abstract

The invention discloses a trajectory planning method based on a multi-stage optimization strategy and a mixed diffusion model. The method comprises the following steps: fusing multi-source sensor data and extracting features; screening and fusing intention anchor point tracks; and optimizing the fine-grained trajectory based on the Transform-Mamba mixed diffusion model. According to the method, an intention anchor point track screening mechanism is provided, the anchor point track matched with the scene is selected from the compact vocabulary through the dynamic screening module and fused with the static prior anchor points, the calculation complexity is reduced, the diversity of the initial track and the scene consistency are guaranteed, and the problem that a traditional fixed anchor point set is insufficient in flexibility is solved. According to the method, a mixed diffusion decoder is designed, space environment dependence is efficiently modeled through a cross attention mechanism, a bidirectional Mama module captures time sequence motion dependence with linear complexity, global perception ability and long sequence modeling efficiency are both considered, and safety and dynamics feasibility of generated tracks are ensured.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, and in particular to a trajectory planning method based on a multi-stage optimization strategy and a hybrid diffusion model. Background Technology

[0002] End-to-end autonomous driving paradigms directly map multi-source sensor data into driving control commands or trajectories using unified deep neural networks, effectively avoiding the error accumulation problem of traditional modular architectures and improving system response efficiency and adaptability to complex scenarios. This has become a research hotspot in the field of autonomous driving. In complex scenarios such as intersections and highway ramp merging, the behavioral intentions of traffic participants are highly uncertain and multimodal. Generating only a single deterministic trajectory can easily lead to planning failure; therefore, multimodal trajectory generation has become a core requirement of end-to-end planning. Existing multimodal trajectory planning methods mainly suffer from the following shortcomings: 1) The method based on a predefined large discrete trajectory vocabulary is essentially a selection of a fixed trajectory set, which limits the model's ability to express diverse driving behaviors and its generation flexibility, and also has high computational complexity. 2) Traditional diffusion models rely on a large number of iterative denoising steps, which are computationally expensive. Meanwhile, the truncation diffusion strategy based on fixed anchor point trajectories lacks scenario adaptability and is difficult to cope with complex and ever-changing traffic environments. 3) A single network structure is difficult to efficiently model the spatial environment dependence and temporal motion dependence required for trajectory planning at the same time, which may lead to collision risks or dynamic infeasibility in the generated trajectory.

[0003] Furthermore, existing methods often lack effective scene adaptation mechanisms in the trajectory initialization phase, making it difficult to provide high-quality initial trajectories for subsequent optimization. They also struggle to balance remote dependency capture and computational efficiency in spatiotemporal modeling. Therefore, how to construct a scene-adaptive initialization mechanism through a multi-stage optimization strategy and design an efficient hybrid network architecture to achieve decoupled modeling of spatiotemporal dependencies has become a key issue in improving the quality, safety, and efficiency of trajectory planning. Summary of the Invention

[0004] To address the aforementioned problems in existing technologies, this invention proposes a trajectory planning method based on a multi-stage optimization strategy and a hybrid diffusion model. This method provides high-quality initialization conditions through intent anchor point trajectory filtering and achieves accurate and efficient multimodal trajectory generation by combining a spatiotemporally decoupled hybrid diffusion model, outputting safe, diverse, and scenario-consistent driving trajectories.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows: a trajectory planning method based on a multi-stage optimization strategy and a hybrid diffusion model, comprising the following steps: A. Fusing multi-source sensor data and extracting features A1. Acquire multi-source sensor data from the autonomous driving system, including the current forward-looking RGB image. LiDAR point cloud data Vehicle status information The vehicle status information Including speed, acceleration, and navigation target; A2. Multi-source sensor data is fused to extract multi-dimensional scene features. Independent feature extraction networks are used to process RGB images and LiDAR point cloud data separately. The Transformer module performs cross-modal feature fusion at each stage to generate image features. and lidar features The LiDAR point cloud is projected onto the bird's-eye view plane, i.e., the BEV plane, and combined with map segmentation information to generate BEV features. After initializing the features of the surrounding agents, the features interact with the BEV features through the Transformer module, and then are decoded by a multilayer perceptron to obtain the features of the surrounding agents. The surrounding intelligent agent features This includes bounding box and category information; an encoder implemented using a multilayer perceptron encodes the vehicle's state information into vehicle state features. The perception module is trained under supervision using bird's-eye view segmentation loss, surrounding agent bounding box loss, and classification loss to ensure the accuracy of feature extraction. The features are represented as follows:

[0006]

[0007]

[0008] In the formula: It is a BEV encoder. It is an encoder for information about surrounding intelligent agents. It is a vehicle status information encoder; These are image features, LiDAR features, BEV features, surrounding intelligent agent features, and vehicle status features; , These are forward-looking RGB image data, LiDAR point cloud data, and vehicle status information; These are the query vector, the query vector initialized with features of surrounding agents, the key vector, and the value vector.

[0009] B. Filter and merge intention anchor point trajectories B1. Construct a compact trajectory vocabulary based on K-means clustering. A fixed-program vocabulary (hereinafter referred to as "vocabulary") is generated from expert trajectories in large-scale human driving data using K-means clustering, and the formula is as follows:

[0010] in , indicating the first A sequence of trajectory points For trajectory indexing, The x-coordinate in the vehicle coordinate system. The vertical coordinate is the coordinate of the vehicle in the vehicle coordinate system. Indicates the first At that moment, The number of trajectories in the vocabulary is represented by '[number]'. A multilayer perceptron is used as the vocabulary encoder to encode each trajectory in the vocabulary as a trajectory feature, ultimately yielding the vocabulary features. The formula is as follows:

[0011] In the formula: It is a vocabulary encoder. It is a vocabulary feature. Represents a vocabulary list.

[0012] B2. Filtering intention anchor trajectories based on attention mechanisms BEV characteristics Vehicle status characteristics By concatenating the features, we obtain the scene context aggregation features. Based on vocabulary features As a query vector and scene context aggregation feature As a key-value vector, the optimized trajectory features are obtained through the Transformer decoder. Then, a score prediction head implemented by a multilayer perceptron calculates the softmax score of each trajectory in the vocabulary; the top-scoring trajector is selected. The trajectory is used as a candidate trajectory, and the formula is as follows:

[0013]

[0014]

[0015]

[0016] In the formula: For splicing operations, For Transformer decoder, The function is softmax, and MLP is a multilayer perceptron. Aggregate features for scene context. The optimized trajectory features; The set of scores for each trajectory in the vocabulary. For the candidate trajectory set, For the candidate trajectory, the first A sequence of trajectory points .

[0017] To prevent the selected trajectory set from being too dense, the spatial distribution of the trajectory is optimized by calculating the root mean square distance between candidate trajectories and employing a farthest point sampling algorithm, ultimately selecting the best trajectories. An intentional anchor trajectory with modal diversity ,in For the intended anchor point trajectory, the first A sequence of trajectory points.

[0018] B3, Fusion Anchor Point Trajectory Introducing N static anchor point trajectories presampled from large-scale human driving data by experts. ,in For the static anchor point trajectory, the first A sequence of trajectory points, this trajectory set encoding rich prior knowledge of driving behavior; the intention anchor point trajectory is fused with the static anchor point trajectory to obtain a hybrid anchor point trajectory set. ,in For the mixed anchor point trajectory, the first A sequence of trajectory points provides high-quality initialization conditions for subsequent trajectory optimization.

[0019] C. Optimize the fine-grained trajectory based on the Transformer-Mamba hybrid diffusion model C1. Initialize the truncation diffusion process First, give the mixed anchor point trajectory set A small amount of Gaussian noise is added to each trajectory to diffuse the mixed anchor point trajectory to an anchored Gaussian distribution, as shown in the following formula:

[0020] In the formula: For diffusion time step index, It is the number of steps to cut off the diffusion. The number of diffusion steps and , It follows a standard Gaussian distribution. For the first Noise figure at each time step For the first The time step The noisy trajectory state of a mixed anchor point trajectory.

[0021] Finally, noisy trajectories are sampled from the anchored Gaussian distribution. A multilayer perceptron is used as the trajectory encoder to encode noisy trajectories into noisy trajectory features. The formula is as follows:

[0022] In the formula: For noisy trajectory features, For trajectory encoder, This is a sinusoidal positional encoding, where D is the feature dimension. C2, Modeling Space Environment Dependency A cross-attention mechanism is used to achieve deep interaction between noisy trajectory features and multi-source environmental information. First, the noisy trajectory features are... Features of BEV Perform deformable spatial cross-attention computation to capture the dependencies between the trajectory and road topology and drivable area; then combine it with features of surrounding agents. Cross-attention calculations are performed to model the interaction between the trajectory and surrounding traffic participants, avoiding collisions; finally, this is combined with the vehicle's state features. Cross-attention calculation is performed to ensure that the trajectory matches the vehicle's driving state, using the following formula:

[0023]

[0024]

[0025] In the formula: Cross-attention in deformable space For agent cross-attention, Cross-pay attention to the vehicle. These are trajectory features resulting from interactions with BEV features, surrounding agent features, and vehicle state features. C3. Modeling temporal motion dependencies After completing the spatial environment interaction, the trajectory features are further interacted with temporal historical information. In continuous driving scenarios, the motion intention should remain continuous and consistent across scenarios. First, the vehicle's historical trajectory is... Encoding as historical trajectory features Then, the spatially enhanced trajectory features are concatenated with historical trajectory features to obtain spatiotemporal fusion features. Finally, a bidirectional Mamba module is introduced to perform forward and reverse scanning of the spatiotemporal fusion features, capturing the temporal continuity of the trajectory with linear complexity, as shown in the following formula:

[0026]

[0027]

[0028] In the formula: This is a bidirectional Mamba module; HT represents the vehicle's historical trajectory. These are historical trajectory features, spatiotemporal fusion features, and trajectory features resulting from interaction with historical trajectory features.

[0029] C4. Decoding and Optimizing Multimodal Trajectories First, a feedforward neural network is used to optimize trajectory features. Then, the time-step modulation layer is used to encode and diffuse the time-step information. Different denoising time steps share parameters; finally, a multilayer perceptron is used to predict the confidence score of each trajectory. and denoising trajectory The trajectory with the highest confidence score is selected as the output, as shown in the following formula:

[0030]

[0031]

[0032] In the formula: FFN is a feedforward neural network. Modulation for time steps; These are trajectory features optimized by a feedforward neural network and trajectory features modulated by time steps, respectively. These are the set of denoised trajectories and the set of confidence scores for denoised trajectories, respectively.

[0033] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention proposes an intent anchor point trajectory filtering mechanism, which selects scene-adaptive anchor point trajectories from a compact vocabulary through a dynamic filtering module and merges them with static prior anchor points. This reduces computational complexity while ensuring the diversity of initial trajectories and scene consistency, thus solving the problem of insufficient flexibility of traditional fixed anchor point sets.

[0034] 2. This invention designs a Transformer-Mamba hybrid diffusion decoder, which efficiently models spatial environment dependencies through a cross-attention mechanism and captures temporal motion dependencies with linear complexity using a bidirectional Mamba module, achieving spatiotemporal decoupling modeling. This balances global perception capabilities with long sequence modeling efficiency, ensuring the safety and dynamic feasibility of the generated trajectory. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating the present invention; Figure 2 This is a framework diagram of a trajectory planning model based on a multi-stage optimization strategy and a hybrid diffusion model. Detailed Implementation

[0036] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0037] like Figure 1 As shown, a trajectory planning method based on a multi-stage optimization strategy and a hybrid diffusion model includes the following steps: A. Feature extraction by fusing multi-source sensor data A1. Acquire multi-source sensor data from the autonomous driving system, including three forward-looking RGB images at the current moment. LiDAR point cloud data Vehicle status information The vehicle status information Including speed, acceleration, and navigation target; A2. Preprocessing and feature fusion of multi-source sensor data. This involves processing three forward-looking RGB images... Cropping and stitching are performed to create a composite image with a resolution of 1024×256×3, which incorporates the LiDAR point cloud. The preprocessed image is a 256×256×1 pseudo-image. Independent ResNet-34 feature extraction networks are used to process the composite image and LiDAR point cloud data separately. A two-layer Transformer decoder is then used to perform cross-modal feature fusion at four scales to generate image features. and lidar features The LiDAR point cloud is projected onto the bird's-eye view plane, i.e., the BEV plane, and combined with map segmentation information to generate BEV features. After initializing the features of the surrounding agents, the features interact with the BEV features through a 3-layer Transformer decoder, and then are decoded by a 2-layer stacked multilayer perceptron to obtain the features of the surrounding agents. The surrounding intelligent agent features This includes bounding box and category information; an encoder implemented using a two-layer stacked multilayer perceptron encodes the vehicle's state information into vehicle state features. The perception module is trained under supervised supervision using bird's-eye view segmentation loss, agent bounding box loss, and classification loss to ensure the accuracy of feature extraction. The features are represented as follows:

[0038]

[0039]

[0040] In the formula: It is a BEV encoder. It is an encoder for information about surrounding intelligent agents. It is a vehicle status information encoder; These are image features, LiDAR features, BEV features, surrounding intelligent agent features, and vehicle status features; , These are forward-looking RGB image data, LiDAR point cloud data, and vehicle status information; These are the query vector, the query vector initialized with features of surrounding agents, the key vector, and the value vector.

[0041] B. Filter and merge intention anchor point trajectories B1. Construct a compact trajectory vocabulary based on K-means clustering. A fixed-program vocabulary (hereinafter referred to as "vocabulary") is generated from expert trajectories in large-scale human driving data using K-means clustering, and the formula is as follows:

[0042] in , indicating the first A sequence of trajectory points For trajectory indexing, The x-coordinate in the vehicle coordinate system. The vertical coordinate in the vehicle coordinate system is denoted as y = x, and each trajectory contains y = x. 8 two-dimensional coordinate points, The number of trajectories in the vocabulary is represented by . A two-layer stacked multilayer perceptron is used as the vocabulary encoder to encode each trajectory in the vocabulary as a trajectory feature, ultimately yielding the vocabulary features. The formula is as follows:

[0043] In the formula: It is a vocabulary encoder; It is a vocabulary feature. Represents a vocabulary list.

[0044] B2. Filtering intent anchor trajectories based on attention mechanisms: BEV characteristics Vehicle status characteristics By concatenating the features, we obtain the scene context aggregation features. Based on vocabulary features As a query vector and scene context aggregation feature As a key-value vector, the optimized trajectory features are obtained through the Transformer decoder. Then, a score prediction head implemented by a three-layer stacked multilayer perceptron calculates the softmax score of each trajectory in the vocabulary; the top-scoring trajector is selected. The trajectory is used as a candidate trajectory, and the formula is as follows:

[0045]

[0046]

[0047]

[0048] In the formula: For splicing operations, For Transformer decoder, The function is softmax, and MLP is a multilayer perceptron. Aggregate features for scene context. The optimized trajectory features; The set of scores for each trajectory in the vocabulary. For the candidate trajectory set, For the candidate trajectory, the first A sequence of trajectory points .

[0049] To prevent the selected trajectory set from being too dense, the spatial distribution of the trajectory is optimized by calculating the root mean square distance between candidate trajectories and employing a farthest point sampling algorithm, ultimately selecting the best trajectories. An intentional anchor trajectory with modal diversity ,in For the intended anchor point trajectory, the first A sequence of trajectory points.

[0050] B3, Fusion Anchor Point Trajectory Introducing N=20 static anchor point trajectories presampled from large-scale human driving data expert trajectories. ,in For the static anchor point trajectory, the first A sequence of trajectory points, this trajectory set encoding rich prior knowledge of driving behavior; the intention anchor point trajectory is fused with the static anchor point trajectory to obtain a hybrid anchor point trajectory set. ,in For the mixed anchor point trajectory, the first A sequence of trajectory points provides high-quality initialization conditions for subsequent trajectory optimization.

[0051] C. Optimize fine-grained trajectories based on the Transformer-Mamba hybrid diffusion decoder. C1. Initialize the truncation diffusion process: First, give the mixed anchor point trajectory set A small amount of Gaussian noise is added to each trajectory to diffuse the mixed anchor point trajectory to an anchored Gaussian distribution, as shown in the following formula:

[0052] In the formula: For diffusion time step index, It is the number of steps to cut off the diffusion. For diffusion steps, It follows a standard Gaussian distribution. For the first Noise figure at each time step For the first The time step The noisy trajectory state of a mixed anchor point trajectory.

[0053] Finally, noisy trajectories are sampled from the anchored Gaussian distribution. A two-layer stacked multilayer perceptron is used as the trajectory encoder to encode the noisy trajectory into noisy trajectory features. The formula is as follows:

[0054] In the formula: For noisy trajectory features, For trajectory encoder, This is a sinusoidal positional encoding, where D is the feature dimension. C2. Modeling Spatial Environment Dependency. A deep interaction between noisy trajectory features and multi-source environmental information is achieved through a cross-attention mechanism. Firstly, the noisy trajectory features... Features of BEV Perform deformable spatial cross-attention computation to capture the dependencies between the trajectory and road topology and drivable area; then combine it with features of surrounding agents. Cross-attention calculations are performed to model the interaction between the trajectory and surrounding traffic participants, avoiding collisions; finally, this is combined with the vehicle's state features. Perform cross-attention calculations to ensure that the trajectory matches the vehicle's driving status.

[0055]

[0056]

[0057]

[0058] In the formula: Cross-attention in deformable space For agent cross-attention, Cross-pay attention to the vehicle. These are trajectory features resulting from interactions with BEV features, surrounding agent features, and vehicle state features. C3. Modeling Temporal Motion Dependencies. After completing the spatial environment interaction, trajectory features further interact with temporal historical information. In continuous driving scenarios, the motion intent should remain continuous and consistent across scenarios. First, the vehicle's historical trajectory is... Encoding as historical trajectory features Then, the spatially enhanced trajectory features are concatenated with historical trajectory features to obtain spatiotemporal fusion features. Finally, a bidirectional Mamba module is introduced to perform forward and reverse scanning of spatiotemporal fusion features, capturing the temporal continuity of the trajectory with linear complexity.

[0059]

[0060]

[0061]

[0062] In the formula: This is a bidirectional Mamba module; HT represents the vehicle's historical trajectory. These are historical trajectory features, spatiotemporal fusion features, and trajectory features resulting from interaction with historical trajectory features.

[0063] C4. Decoding and Optimizing Multimodal Trajectories: DDIM sampling is used, with a sampling step count of 2. First, a feedforward neural network is employed to optimize the trajectory features. Then, the time-step modulation layer is used to encode and diffuse the time-step information. Different denoising time steps share parameters; finally, a multilayer perceptron is used to predict the confidence score of each trajectory. and denoising trajectory The trajectory with the highest confidence score is selected as the output.

[0064]

[0065]

[0066]

[0067] In the formula: FFN is a feedforward neural network. Modulation for time steps; These are trajectory features optimized by a feedforward neural network and trajectory features modulated by time steps, respectively. These are the set of denoised trajectories and the set of confidence scores for denoised trajectories, respectively.

[0068] The basic principles, main features, and advantages of this invention have been described above. Those skilled in the art should understand that this invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of this invention. Various changes and modifications can be made to this invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the invention as claimed.

Claims

1. A trajectory planning method based on a multi-stage optimization strategy and a hybrid diffusion model, characterized in that: Includes the following steps: A. Integrate multi-source sensor data and extract features; B. Filter and merge intent anchor point trajectories; C. Optimize fine-grained trajectories based on the Transformer-Mamba hybrid diffusion model.

2. The trajectory planning method based on a multi-stage optimization strategy and a hybrid diffusion model according to claim 1, characterized in that: The method for fusing multi-source sensor data and extracting features as described in step A includes the following steps: A1. Acquire multi-source sensor data from the autonomous driving system, including the current forward-looking RGB image. LiDAR point cloud data Vehicle status information The vehicle status information Including speed, acceleration, and navigation target; A2. Multi-source sensor data is fused to extract multi-dimensional scene features. Independent feature extraction networks are used to process RGB images and LiDAR point cloud data separately. The Transformer module performs cross-modal feature fusion at each stage to generate image features. and lidar features The LiDAR point cloud is projected onto the bird's-eye view plane, i.e., the BEV plane, and combined with map segmentation information to generate BEV features. After initializing the features of the surrounding agents, the features interact with the BEV features through the Transformer module, and then are decoded by a multilayer perceptron to obtain the features of the surrounding agents. The surrounding intelligent agent features This includes bounding box and category information; an encoder implemented using a multilayer perceptron encodes the vehicle's state information into vehicle state features. The perception module is trained under supervision using bird's-eye view segmentation loss, surrounding agent bounding box loss, and classification loss to ensure the accuracy of feature extraction. The features are represented as follows: In the formula: It is a BEV encoder. It is an encoder for information about surrounding intelligent agents. It is a vehicle status information encoder; These are image features, LiDAR features, BEV features, surrounding intelligent agent features, and vehicle status features; , These are forward-looking RGB image data, LiDAR point cloud data, and vehicle status information; These are the query vector, the query vector initialized with features of surrounding agents, the key vector, and the value vector.

3. The trajectory planning method based on a multi-stage optimization strategy and a hybrid diffusion model according to claim 1, characterized in that: Step B, the method for filtering and fusing intent anchor point trajectories, includes the following steps: B1. Construct a compact trajectory vocabulary based on K-means clustering. A fixed-program vocabulary (hereinafter referred to as "vocabulary") is generated from expert trajectories in large-scale human driving data using K-means clustering, and the formula is as follows: in , indicating the first A sequence of trajectory points For trajectory indexing, The x-coordinate in the vehicle coordinate system. The vertical coordinate is the coordinate of the vehicle in the vehicle coordinate system. Indicates the first At that moment, The number of trajectories in the vocabulary is given; a multilayer perceptron is used as the vocabulary encoder to encode each trajectory in the vocabulary as a trajectory feature, ultimately obtaining the vocabulary features. The formula is as follows: In the formula: It is a vocabulary encoder. It is a vocabulary feature. Represents a vocabulary list; B2. Filtering intention anchor trajectories based on attention mechanisms BEV characteristics Vehicle status characteristics By concatenating the features, we obtain the scene context aggregation features. Based on vocabulary features As a query vector and scene context aggregation feature As a key-value vector, the optimized trajectory features are obtained through the Transformer decoder. Then, a score prediction head implemented by a multilayer perceptron calculates the softmax score of each trajectory in the vocabulary; the top-scoring trajector is selected. The trajectory is used as a candidate trajectory, and the formula is as follows: In the formula: For splicing operations, For Transformer decoder, The function is softmax, and MLP is a multilayer perceptron. Aggregate features for scene context. The optimized trajectory features; The set of scores for each trajectory in the vocabulary. For the candidate trajectory set, For the candidate trajectory, the first A sequence of trajectory points ; To prevent the selected trajectory set from being too dense, the spatial distribution of the trajectory is optimized by calculating the root mean square distance between candidate trajectories and employing a farthest point sampling algorithm, ultimately selecting the best trajectories. An intentional anchor trajectory with modal diversity ,in For the intended anchor point trajectory, the first A sequence of trajectory points; B3, Fusion Anchor Point Trajectory Introducing N static anchor point trajectories presampled from large-scale human driving data by experts. ,in For the static anchor point trajectory, the first A sequence of trajectory points, this trajectory set encoding rich prior knowledge of driving behavior; the intention anchor point trajectory is fused with the static anchor point trajectory to obtain a hybrid anchor point trajectory set. ,in For the mixed anchor point trajectory, the first A sequence of trajectory points provides high-quality initialization conditions for subsequent trajectory optimization.

4. The trajectory planning method based on a multi-stage optimization strategy and a hybrid diffusion model according to claim 1, characterized in that: Step C describes the method for optimizing fine-grained trajectories based on the Transformer-Mamba hybrid diffusion model, which includes the following steps: C1. Initialize the truncation diffusion process First, give the mixed anchor point trajectory set A small amount of Gaussian noise is added to each trajectory to diffuse the mixed anchor point trajectory to an anchored Gaussian distribution, as shown in the following formula: In the formula: For diffusion time step index, It is the number of steps to cut off the diffusion. The number of diffusion steps and , It follows a standard Gaussian distribution. For the first Noise figure at each time step For the first The time step The noisy trajectory state of a mixed anchor point trajectory; Finally, noisy trajectories are sampled from the anchored Gaussian distribution. A multilayer perceptron is used as the trajectory encoder to encode noisy trajectories into noisy trajectory features. The formula is as follows: In the formula: For noisy trajectory features, For trajectory encoder, This is a sinusoidal positional encoding, where D is the feature dimension. C2, Modeling Space Environment Dependency A cross-attention mechanism is used to achieve deep interaction between noisy trajectory features and multi-source environmental information. First, the noisy trajectory features are... Features of BEV Perform deformable spatial cross-attention computation to capture the dependencies between the trajectory and road topology and drivable area; then combine it with features of surrounding agents. Cross-attention calculations are performed to model the interaction between the trajectory and surrounding traffic participants, avoiding collisions; finally, this is combined with the vehicle's state features. Cross-attention calculation is performed to ensure that the trajectory matches the vehicle's driving state, using the following formula: In the formula: Cross-attention in deformable space For agent cross-attention, Cross-attention between vehicles; These are trajectory features resulting from interactions with BEV features, surrounding agent features, and vehicle state features. C3. Modeling temporal motion dependencies After completing the spatial environment interaction, the trajectory features are further interacted with temporal historical information. In continuous driving scenarios, the motion intention should remain continuous and consistent across scenarios. First, the vehicle's historical trajectory is... Encoding as historical trajectory features Then, the spatially enhanced trajectory features are concatenated with historical trajectory features to obtain spatiotemporal fusion features. Finally, a bidirectional Mamba module is introduced to perform forward and reverse scanning of the spatiotemporal fusion features, capturing the temporal continuity of the trajectory with linear complexity, as shown in the following formula: In the formula: This is a bidirectional Mamba module; HT represents the vehicle's historical trajectory. These are historical trajectory features, spatiotemporal fusion features, and trajectory features resulting from interaction with historical trajectory features. C4. Decoding and Optimizing Multimodal Trajectories First, a feedforward neural network is used to optimize trajectory features. Then, the time-step modulation layer is used to encode and diffuse the time-step information. Different denoising time steps share parameters; finally, a multilayer perceptron is used to predict the confidence score of each trajectory. and denoising trajectory The trajectory with the highest confidence score is selected as the output, as shown in the following formula: In the formula: FFN is a feedforward neural network. Modulation for time steps; These are trajectory features optimized by a feedforward neural network and trajectory features modulated by time steps, respectively. These are the set of denoised trajectories and the set of confidence scores for denoised trajectories, respectively.

Images