Trajectory prediction method and system based on gated attention and probabilistic multi-modal

Abstract

This invention discloses a trajectory prediction method and system based on gated attention and probabilistic multimodal analysis, relating to the field of autonomous driving technology. The method includes: acquiring and encoding historical trajectory data of multiple vehicles in a target scene; inputting the encoded features into an interaction enhancement module, reweighting the multi-head self-attention weights by introducing a dynamic relationship-based gating calibration factor to refine the modeling of vehicle interactions; subsequently, inputting the enhanced features into a multimodal prediction head, which is based on a conditional variational autoencoder architecture, generating multiple future trajectories and outputting an independent explicit probability value for each trajectory through a parallel probability estimation branch. This invention, through a gating calibration attention mechanism and a multimodal generation structure capable of outputting explicit probabilities, solves the problems of insufficient fine-grained interaction modeling and ambiguous multimodal prediction probabilities in existing technologies, significantly improving the accuracy, diversity, and interpretability of trajectory prediction.

Description

Technical Field

[0001] This invention relates to the fields of autonomous driving and artificial intelligence technology, and in particular to a trajectory prediction method and system based on gated attention and probabilistic multimodal approaches. Background Technology

[0002] Accurately predicting the future trajectories of surrounding vehicles is a crucial prerequisite for safe navigation and decision-making in autonomous driving systems. The core challenge of this task lies in simultaneously addressing two aspects: first, complex interaction modeling, which involves understanding the dynamic and mutually influential relationships between multiple vehicles; and second, multimodal prediction, which requires accurate and quantifiable predictions of various plausible future possibilities.

[0003] Currently, deep learning-based trajectory prediction methods have made significant progress. Among them, attention-based models are widely used due to their powerful relationship modeling capabilities. However, existing technologies still have the following shortcomings: First, in terms of interaction modeling, most methods adopt standard attention mechanisms, whose attention weights are only calculated based on feature similarity, failing to explicitly incorporate the dynamic kinematic relationships between vehicles (such as relative speed and distance), resulting in insufficient refinement in characterizing the importance of interactions. Second, in terms of multimodal prediction, mainstream methods such as conditional variational autoencoders (CVAEs) can generate diverse trajectories, but they usually cannot directly output explicit, interpretable probability values ​​for each trajectory, or require complex post-processing estimation, limiting the direct application of prediction results in safety-critical decisions.

[0004] Therefore, there is an urgent need for a prediction method that can perform more refined modeling of vehicle interactions and directly output multimodal trajectories with explicit probability evaluation, so as to improve the accuracy, interpretability and reliability of predictions. Summary of the Invention

[0005] To address the above technical problems, this invention provides a trajectory prediction method based on gated attention and probabilistic multimodal analysis, comprising the following steps: Step S1: Obtain the historical trajectory sequence of the vehicle and multiple intelligent agents around it in the target scene. The intelligent agents include vehicles and pedestrians. Step S2: Encode the historical trajectory sequence to obtain the temporal trajectory features of the vehicle and multiple intelligent agents around it; Step S3: Input the temporal trajectory features into the interaction enhancement module for processing to obtain interaction enhancement features; The interaction enhancement module introduces a gating calibration factor based on the relative motion state between the vehicle and multiple intelligent agents around it, and dynamically reweights the initial weights of the standard multi-head self-attention calculation. The weighted aggregated features are then passed through a feedforward network to output the interaction enhancement features. Step S4: Input the interactive enhancement features into the multimodal prediction head for processing to generate multiple modal predictions of future trajectories and corresponding explicit probability distributions; the multimodal prediction head is built based on the conditional variational autoencoder architecture and adds a probability estimation branch that runs in parallel with the decoder. The explicit probability value of each generated future trajectory is directly calculated through the probability estimation branch. Step S5: Output the multimodal prediction results, which include K future trajectories and their corresponding explicit probability values.

[0006] The technical solution further defined in this invention is: Furthermore, in step S1, the vehicle's sensors and perception algorithms acquire data on the vehicle and multiple intelligent agents in its surroundings in the past. The historical trajectory coordinates at each time step are then normalized to form a historical trajectory sequence. Where N represents the number of agents. The historical time step represents the number of frames observed in retrospect, i.e., the duration of the time.

[0007] As described above, the trajectory prediction method based on gated attention and probabilistic multimodal approaches involves, in step S2, using a Transformer encoder or graph neural network to encode the historical trajectory sequence X, extracting features of the vehicle and each agent around it at each moment, and fusing spatiotemporal information to output temporal trajectory features. N represents the number of vehicles and intelligent agents, and D represents the feature dimension.

[0008] As described above, the trajectory prediction method based on gated attention and probabilistic multimodal approaches includes the following sub-steps in step S3: Step S3.1: Calculate the initial attention weights between each pair of the vehicle and its surrounding agents using a multi-head self-attention mechanism; Step S3.2: Introduce a gating calibration factor based on the relative motion state between the vehicle and multiple intelligent agents around it, and dynamically reweight the initial attention weights to obtain the calibrated attention weights; Step S3.3: Use the calibrated attention weights to perform weighted aggregation of temporal trajectory features and output interactive enhancement features.

[0009] As described above, in the trajectory prediction method based on gated attention and probabilistic multimodal, in step S3.2, the gated calibration factor is generated by a multilayer perceptron. The input of the multilayer perceptron is the relative displacement vector, relative velocity vector and historical distance between the vehicle and multiple intelligent agents around it; and the gated calibration factor is multiplied by the initial attention weight.

[0010] As described above, the trajectory prediction method based on gated attention and probabilistic multimodal approaches includes the following sub-steps in step S4: Step S4.1: Using the interactive enhancement features as conditions, sample K different latent variables from the prior distribution; Step S4.2: Concatenate each latent variable with the conditions and generate K future trajectories using the decoder; Step S4.3: Calculate an independent explicit probability value for each of the K future trajectories through a probability estimation branch that runs in parallel with the decoder network.

[0011] As described above, in the trajectory prediction method based on gated attention and probabilistic multimodal, in step S4.3, the probability estimation branch is set as an independent multilayer perceptron, which receives the features of the middle layer of the decoder as input while the decoder generates each future trajectory, and outputs a scalar probability value; the scalar probability values ​​of all K future trajectories are normalized by the Softmax function to form a discrete probability distribution.

[0012] As described above, in the trajectory prediction method based on gated attention and probabilistic multimodal approaches, after obtaining the temporal trajectory features in step S2, when there is a digital map containing lane lines and traffic sign semantic information, and with an absolute positioning accuracy better than 10 cm and a relative accuracy better than 5 cm, the temporal trajectory features of the vehicle and each intelligent agent around it are fused with the map semantic features of the lane segment in which it is located through the multi-layer graph attention mechanism of the dynamic heterogeneous graph attention network to generate spatiotemporal fusion features, and the original temporal trajectory features are replaced with spatiotemporal fusion features and input into step S3.

[0013] As described above, the trajectory prediction method based on gated attention and probabilistic multimodal approaches uses a dynamic heterogeneous graph attention network. In this network, the vehicle and its surrounding agents and lane segments are modeled as two types of nodes. Edges are constructed based on the relationships between the vehicle and its surrounding agents and lane segments, as well as the connections between lane segments. The graph attention mechanism is used to propagate and fuse features across node types, outputting spatiotemporal features that incorporate scene constraints, ensuring that the generated trajectory conforms to the road topology.

[0014] A trajectory prediction system based on gated attention and probabilistic multimodal approaches, including... The data acquisition module is used to acquire the historical trajectory sequences of the vehicle and multiple intelligent agents around it in the target scene; The feature encoding module is used to encode the historical trajectory sequence to obtain the temporal trajectory features of the vehicle and each intelligent agent around it; The interaction enhancement module is used to process the temporal trajectory features to obtain interaction enhancement features; and introduces a gating calibration factor based on the relative motion state between the vehicle and multiple intelligent agents around it to dynamically reweight the initial weights of the standard multi-head self-attention calculation. The weighted aggregated features are then passed through a feedforward network to output the interaction enhancement features. The multimodal prediction module is used to input interactive enhancement features into the multimodal prediction head for processing, generating multiple modal predictions of future trajectories and corresponding explicit probability distributions. The multimodal prediction head is built on a conditional variational autoencoder architecture and adds a probability estimation branch that runs in parallel with the decoder. The explicit probability value of each generated future trajectory is directly calculated through the probability estimation branch. The prediction output interface module is used to package multiple sets of formatted future trajectories and their explicit probability values, and send them to the vehicle's decision-making and path planning module.

[0015] The beneficial effects of this invention are: (1) The interaction modeling in this invention is more refined. By introducing a gating calibration factor based on dynamic motion relationship, the attention mechanism can better focus on vehicle pairs with potential collision risks or strong mutual influence, thereby improving the feature representation capability in complex interaction scenarios. (2) In this invention, the probability output is more explicit. Through an independent probability estimation branch, the normalized probability of each trajectory is directly output while generating multimodal trajectories. No post-processing is required, which enhances the interpretability of the prediction results and facilitates risk quantification assessment by downstream decision-making modules. (3) The scene fusion in this invention is more effective. The deep fusion of trajectory and map features is achieved through dynamic heterogeneous graph attention network, so that the predicted trajectory strictly follows the lane-level road structure, which improves the rationality and accuracy of prediction in structured road environment. (4) The system integration in this invention is higher, the whole method can be trained end-to-end, the modular design is clear, and it is easy to implement and deploy on the autonomous driving computing platform. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 This is a schematic diagram of the workflow of the interaction enhancement module in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the workflow of the multimodal prediction head in an embodiment of the present invention. Detailed Implementation

[0017] This embodiment provides a trajectory prediction method based on gated attention and probabilistic multimodal approaches, such as... Figure 1 As shown, it includes the following steps: Step S1: Obtain information from vehicle-mounted sensors (cameras, LiDAR) and perception algorithms regarding multiple intelligent agents (vehicles, pedestrians, etc.) around the vehicle in the target scene in the past. The historical trajectory coordinates are obtained at each time step (e.g., 2 seconds, 10 frames per second); and the coordinates are normalized to form a historical trajectory sequence. Where N represents the number of agents. The historical time step represents the number of frames (duration) observed in the retrospective view.

[0018] Step S2: Encode the historical trajectory sequence to obtain the temporal trajectory features of each agent.

[0019] The historical trajectory sequence X is encoded using a Transformer encoder or graph neural network. This encoder extracts features from each agent at each time step and initially fuses spatiotemporal information to output time-series trajectory features. N represents the number of agents, and D represents the feature dimension.

[0020] To utilize the prior information of high-precision maps, when a high-precision map exists (a high-precision map refers to a digital map that includes lane lines and traffic sign semantic information, and has an absolute positioning accuracy better than 10 cm and a relative accuracy better than 5 cm), after obtaining the temporal trajectory features in step S2, a dynamic heterogeneous graph attention network is used to deeply fuse the temporal trajectory features of each agent with the map semantic features of its lane segment to generate spatiotemporal fusion features, and the original temporal trajectory features are replaced with spatiotemporal fusion features and input into step S3.

[0021] In the Dynamic Heterogeneous Graph Attention Network, a heterogeneous graph is constructed, containing two types of heterogeneous nodes: vehicle nodes and lane segment nodes. Edges between nodes are dynamically constructed based on the actual spatial affiliation of vehicles and lane segments (e.g., a vehicle is located within a lane segment) and the topological connections between lane segments (e.g., connectivity). This network utilizes a multi-layer graph attention mechanism to perform bidirectional feature message passing and aggregation between vehicle nodes and lane segment nodes, thereby achieving deep fusion of trajectory features and map semantic features (e.g., lane type, curvature, traffic rule constraints), outputting a spatiotemporal fusion feature F_fused for subsequent steps.

[0022] In the dynamic heterogeneous graph attention network, agents and lane segments are modeled as two types of nodes respectively. Edges are constructed based on the relationship between agents and lane segments and the connection relationship between lane segments. The graph attention mechanism is used to propagate and fuse features across node types, and outputs spatiotemporal features that incorporate scene constraints to ensure that the generated trajectory conforms to the road topology.

[0023] Step S3: Input the temporal trajectory feature F_traj into the interaction enhancement module for processing to obtain the interaction enhancement feature; the interaction enhancement module introduces a gating calibration factor based on the relative motion state between agents to dynamically reweight the initial weights of the standard multi-head self-attention calculation, thereby achieving refined interaction modeling. The weighted and aggregated features are then passed through a feedforward network to output the interaction enhancement feature.

[0024] like Figure 2As shown, the core of the interaction enhancement module is a gated calibration multi-head self-attention layer. Its workflow is as follows: First, the initial attention weights between agents are calculated; then, a gated generation sub-network calculates a scalar gate calibration factor based on dynamic features such as relative displacement and relative velocity between paired agents, and reweights the initial attention weights to enhance key interactions and suppress irrelevant interference; finally, the weighted aggregated features are passed through a feedforward network to output interaction enhancement features. , where N is the number of agents and D is the feature dimension.

[0025] Step S3 specifically includes the following sub-steps: Step S3.1: Calculate the initial attention weights between all agents in pairs using a multi-head self-attention mechanism.

[0026] Step S3.2: Introduce a gating calibration factor based on the relative motion state between agents to dynamically reweight the initial attention weights and obtain calibrated attention weights. The gating calibration factor is generated by a lightweight neural network (such as a multilayer perceptron). The input of the multilayer perceptron is the relative motion features between agents (displacement vector, relative velocity vector, and historical distance). The calibration factor is multiplied by the initial attention weights to amplify the weights of key interactions (such as rapidly approaching vehicle pairs) and suppress the influence of irrelevant or weak interactions, thereby outputting enhanced features that better reflect the real driving interaction mode.

[0027] Step S3.3: Use the calibrated attention weights to perform weighted aggregation of temporal trajectory features and output interactive enhancement features.

[0028] Step S4: Input the interactive enhancement feature F_enhanced into the multimodal prediction head for processing, generating multiple modal predictions of the future trajectory and the corresponding explicit probability distributions; such as Figure 3 As shown, the multimodal prediction head is built on a conditional variational autoencoder (CVAE) architecture and adds a probability estimation branch that runs in parallel with the decoder. The explicit probability value of each generated future trajectory is directly calculated through the probability estimation branch.

[0029] Step S4 specifically includes the following sub-steps: Step S4.1: Using the interactive enhancement feature F_enhanced as a condition, sample K different latent variables from the prior distribution N(0,I). .

[0030] Step S4.2: For each latent variable... Each of these conditions is concatenated with the condition F_enhanced and input into the main decoder to generate a future trajectory. .

[0031] Step S4.3: Calculate an independent explicit probability value for each of the K future trajectories through a probability estimation branch that runs in parallel with the decoder network.

[0032] The probability estimation branch is configured as a separate multilayer perceptron, which receives features from the middle layers of the decoder as input while the decoder generates each future trajectory; the same conditional latent variables are input to a parallel probability estimation branch (another multilayer perceptron), which outputs a scalar probability value. ; Put all K scalars By normalizing using the Softmax function to represent the confidence distribution of each prediction mode, the likelihood of each predicted trajectory is intuitively reflected, yielding the explicit probability value corresponding to each trajectory. This forms a discrete probability distribution, where e represents the natural exponent; ultimately, K sets of trajectory-probability pairs are obtained. ,in , For future time steps.

[0033] Step S5: Output the multimodal prediction results, which include K future trajectories and their corresponding explicit probability values.

[0034] In this embodiment, it is assumed that a vehicle with L4 autonomous driving capability (the autonomous vehicle) is entering an intersection without traffic lights. It is necessary to predict the future trajectories of three interacting vehicles around it (denoted as V1, V2, and V3) in order to carry out safety planning.

[0035] First, obtain N=4 vehicles: V1, V2, and V3. The trajectory coordinates within a second are sampled at a frequency of 10Hz, meaning there are a total of 30 historical frames. The historical trajectory sequence is represented as follows: .

[0036] Then load a centimeter-level high-precision map of the intersection (absolute accuracy better than 10 centimeters), which includes vectorized semantic information such as lane center lines, boundaries, and stop lines.

[0037] Next, feature encoding is performed: X is encoded using a Transformer encoder. This encoder consists of three stacked encoding layers, outputting trajectory features. The feature dimension D is 128.

[0038] Then, spatiotemporal feature fusion is performed: based on the high-precision map, the current lane segment of each vehicle is determined. A dynamic heterogeneous graph containing 4 vehicle nodes and several lane segment nodes is constructed. Feature propagation and fusion are performed through a 2-layer graph attention network, outputting spatiotemporal fused features. .

[0039] Then, interaction enhancement is performed: F_fused is input into the interaction enhancement module. The core of this module is a gated calibration multi-head self-attention layer (with 8 heads). The gated generative subnetwork (a two-layer MLP) calculates the calibration factor based on the relative displacement and velocity between vehicles, dynamically adjusts the attention weights, and outputs the interaction enhancement feature F_enhanced.

[0040] Then, multimodal prediction is performed: F_enhanced is input into the multimodal prediction module. This module contains a backbone of a conditional variational autoencoder and a parallel probability estimation branch (a 3-layer MLP).

[0041] In multimodal prediction, K=5 latent variables are sampled from the prior distribution; The main decoder (a 4-layer MLP) generates the future for each variable. Tracking per second (50 frames) Meanwhile, the probability estimation branch generates a scalar value for each trajectory, which is then normalized using Softmax to obtain five explicit probability values. (k=1 to 5).

[0042] Finally, the system outputs the results. In this embodiment, the system ultimately outputs five sets of possible future trajectories and their corresponding probabilities: 1. Trajectory 1 (probability) V1 slows down to give way, V2 accelerates to go straight through, V3 turns right.

[0043] 2. Trajectory 2 (probability) ): V1 passes through at a constant speed, V2 brakes slightly, and V3 turns right.

[0044] 3. Trajectory 3 (probability) V1 is for emergency braking, V2 is for emergency avoidance, and V3 remains unchanged.

[0045] 4. Trajectory 4 (probability) V1 and V2 pass through at a constant speed, V3 turns left.

[0046] 5. Trajectory 5 (probability) ): Other low-probability behavior combinations.

[0047] These multimodal prediction results have explicit probability values. It will be transmitted in real time to the downstream trajectory planning and decision-making module, serving as a key basis for risk assessment and the generation of safe driving trajectories.

[0048] To verify the effectiveness of the trajectory prediction method based on gated calibrated attention and explicit probabilistic multimodal generation proposed in this embodiment, comprehensive experiments were conducted on the publicly available pedestrian interaction dataset ETH / UCY. This dataset includes five scenarios: eth, hotel, univ, zara1, and zara2, and serves as a benchmark for evaluating multi-agent trajectory prediction models.

[0049] Evaluation metrics: The average displacement error (ADE↓) and final displacement error (FDE↓), commonly used in trajectory prediction, are adopted as evaluation metrics, where "↓" indicates that the lower the value, the better. ADE measures the average error of the entire predicted trajectory, while FDE measures the error at the trajectory's endpoint.

[0050] Baseline Model: The classic Transformer-based multi-agent trajectory prediction model AgentFormer was selected as the strong baseline.

[0051] Comparison Setup: To verify the independent contributions and synergistic effects of the two core modules in this invention, the following ablation experimental models were set up for comparison:

[0052] Baseline: The original AgentFormer model.

[0053] Baseline+IA (Interactive Enhancement Module): This adds only the gating calibration interactive enhancement module of this embodiment to the Baseline.

[0054] Baseline+CVAE (Multimodal Prediction Header): This adds the CVAE multimodal prediction head with parallel probability estimation branch of this embodiment to the Baseline.

[0055] Ours (Model of this embodiment): A complete model that integrates all the innovative aspects of the method in this embodiment.

[0056] Table 1

[0057] As shown in Table 1, the results were analyzed: 1. The model in this embodiment has the best performance: The model in this embodiment (Ours) outperforms the baseline model (AgentFormer) in all aspects of the ETH scenario and the average metrics of the five scenarios. For example, in the more difficult ETH scenario, ADE and FDE are reduced by 24.4% and 26.7% respectively, significantly improving prediction accuracy.

[0058] 2. Module-independent effectiveness: Models with either the "Gated Calibration Interaction Enhancement Module (IA)" or the "Explicit Probabilistic CVAE Prediction Head (CVAE)" added individually outperformed the original baseline. This demonstrates the effectiveness of both refined interaction modeling and explicit probabilistic multimodal generation techniques.

[0059] 3. Synergistic effect exists between modules: The performance improvement of the model (Ours) in this embodiment is greater than the simple sum of the performance improvements of the two independent improvement modules (Baseline+IA, Baseline+CVAE). This indicates that the two core modules in this embodiment do not work in isolation, but provide better conditional inputs for subsequent multimodal probabilistic generation through the interactive features of gating calibration. The two work together to produce a "1+1>2" effect.

[0060] In summary, the experimental data fully validates the effectiveness of the "gated calibration attention mechanism" and "parallel explicit probability estimation branch" proposed in this embodiment in improving the accuracy of trajectory prediction, as well as the good synergy between modules, supporting the innovativeness and practicality claims of this invention.

[0061] This embodiment also provides a trajectory prediction system based on gated attention and probabilistic multimodal analysis, deployed on an autonomous driving computing platform. The system includes a sensor and data interface module, a perception and data preparation module, a feature calculation module, a multimodal prediction module, and a prediction output interface module. These modules work collaboratively to achieve end-to-end processing from raw data to probabilistic multimodal trajectory prediction.

[0062] Sensor and data interface module: used to receive raw perception data from vehicle cameras and LiDAR, and to access the vehicle's high-precision map data.

[0063] Perception and Data Preparation Module: Connected to the sensor and data interface module, it is used to perform target detection and tracking on raw perception data, output historical trajectory sequences of each vehicle in the environment, and parse high-precision map data to extract lane-level semantic information.

[0064] Feature computation module: Connected to the perception and data preparation module, it contains the hardware acceleration core of the encoder, spatiotemporal feature fusion submodule, and interaction enhancement submodule; the encoder is used to encode features of historical trajectory sequences; the spatiotemporal feature fusion submodule is configured to fuse trajectory features and map semantic features through a dynamic heterogeneous graph attention network; the interaction enhancement submodule is configured to perform refined interaction modeling of features using a gated calibration multi-head self-attention mechanism.

[0065] Multimodal prediction module: Connected to the feature calculation module, it contains the hardware acceleration core of the multimodal prediction head; the multimodal prediction head is based on the conditional variational autoencoder architecture and integrates a parallel probability estimation branch, which is used to synchronously generate multiple future trajectories and their corresponding explicit probability values ​​based on interactive enhancement features.

[0066] Prediction output interface module: Connected to the multimodal prediction module, it is used to package the formatted multiple sets of trajectories and their probabilities and send them to the vehicle's decision and path planning module.

[0067] This embodiment addresses the issues of insufficient fine-grained interaction modeling and unclear multimodal prediction probabilities in existing technologies by employing a gating calibration attention mechanism and a multimodal generation structure capable of outputting explicit probabilities. This significantly improves the accuracy, diversity, and interpretability of trajectory prediction.

[0068] In addition to the embodiments described above, the present invention may have other implementations. All technical solutions formed by equivalent substitution or equivalent transformation fall within the protection scope claimed by the present invention.

Claims

1. A trajectory prediction method based on gated attention and probabilistic multimodal analysis, characterized in that: Includes the following steps: Step S1: Obtain the historical trajectory sequence of the vehicle and multiple intelligent agents around it in the target scene. The intelligent agents include vehicles and pedestrians. Step S2: Encode the historical trajectory sequence to obtain the temporal trajectory features of the vehicle and multiple intelligent agents around it; Step S3: Input the temporal trajectory features into the interaction enhancement module for processing to obtain the interaction enhancement features; The interaction enhancement module introduces a gating calibration factor based on the relative motion state between the vehicle and multiple intelligent agents around it, dynamically reweights the initial weights of the standard multi-head self-attention calculation, and outputs the interaction enhancement features through a feedforward network after weighted aggregation. Step S4: Input the interactive enhancement features into the multimodal prediction head for processing to generate multiple modal predictions of future trajectories and corresponding explicit probability distributions; the multimodal prediction head is built based on the conditional variational autoencoder architecture and adds a probability estimation branch that runs in parallel with the decoder. The explicit probability value of each generated future trajectory is directly calculated through the probability estimation branch. Step S5: Output the multimodal prediction results, which include K future trajectories and their corresponding explicit probability values.

2. The trajectory prediction method based on gated attention and probabilistic multimodal analysis according to claim 1, characterized in that: In step S1, the vehicle sensors and perception algorithms acquire information about the vehicle and multiple intelligent agents in the past. The historical trajectory coordinates at each time step are then normalized to form a historical trajectory sequence. Where R represents the input data X as a mathematical object consisting of real numbers, and N represents the number of agents. The historical time step represents the number of frames observed in retrospect, i.e., the duration of the time.

3. The trajectory prediction method based on gated attention and probabilistic multimodal analysis according to claim 2, characterized in that: In step S2, a Transformer encoder or graph neural network is used to encode the historical trajectory sequence X, extracting features of the vehicle and each agent around it at each moment, and fusing spatiotemporal information to output temporal trajectory features. N represents the number of vehicles and intelligent agents, and D represents the feature dimension.

4. The trajectory prediction method based on gated attention and probabilistic multimodal analysis according to claim 1, characterized in that: Step S3 specifically includes the following sub-steps: Step S3.1: Calculate the initial attention weights between each pair of the vehicle and its surrounding agents using a multi-head self-attention mechanism; Step S3.2: Introduce a gating calibration factor based on the relative motion state between the vehicle and multiple intelligent agents around it, and dynamically reweight the initial attention weights to obtain the calibrated attention weights; Step S3.3: Use the calibrated attention weights to perform weighted aggregation of temporal trajectory features and output interactive enhancement features.

5. The trajectory prediction method based on gated attention and probabilistic multimodal analysis according to claim 4, characterized in that: In step S3.2, the gating calibration factor is generated by a multilayer perceptron. The input of the multilayer perceptron is the relative displacement vector, relative velocity vector and historical distance between the vehicle and multiple intelligent agents around it; and the gating calibration factor is multiplied by the initial attention weight.

6. The trajectory prediction method based on gated attention and probabilistic multimodal analysis according to claim 1, characterized in that: Step S4 specifically includes the following sub-steps: Step S4.1: Using the interactive enhancement features as conditions, sample K different latent variables from the prior distribution; Step S4.2: Concatenate each latent variable with the conditions and generate K future trajectories using the decoder; Step S4.3: Calculate an independent explicit probability value for each of the K future trajectories through a probability estimation branch that runs in parallel with the decoder network.

7. The trajectory prediction method based on gated attention and probabilistic multimodal analysis according to claim 6, characterized in that: In step S4.3, the probability estimation branch is set as an independent multilayer perceptron, which receives the features of the middle layer of the decoder as input while the decoder generates each future trajectory, and outputs a scalar probability value; the scalar probability values ​​of all K future trajectories are normalized by the Softmax function to form a discrete probability distribution.

8. The trajectory prediction method based on gated attention and probabilistic multimodal analysis according to claim 1, characterized in that: After obtaining the temporal trajectory features in step S2, when there is a digital map containing lane lines and traffic sign semantic information, and with an absolute positioning accuracy better than 10 cm and a relative accuracy better than 5 cm, the temporal trajectory features of the vehicle and each intelligent agent around it are fused with the map semantic features of the lane segment in which it is located through the multi-layer graph attention mechanism of the dynamic heterogeneous graph attention network to generate spatiotemporal fusion features, and the original temporal trajectory features are replaced with spatiotemporal fusion features and input into step S3.

9. The trajectory prediction method based on gated attention and probabilistic multimodal analysis according to claim 8, characterized in that: In the dynamic heterogeneous graph attention network, the vehicle and its surrounding agents and lane segments are modeled as two types of nodes respectively. Edges are constructed based on the relationship between the vehicle and its surrounding agents and lane segments, as well as the connection relationship between lane segments. The graph attention mechanism is used to propagate and fuse features across node types, and outputs spatiotemporal features that incorporate scene constraints to ensure that the generated trajectory conforms to the road topology.

10. A trajectory prediction system based on gated attention and probabilistic multimodal analysis, used to perform the method described in any one of claims 1-9, characterized in that: include The data acquisition module is used to acquire the historical trajectory sequences of the vehicle and multiple intelligent agents around it in the target scene; The feature encoding module is used to encode the historical trajectory sequence to obtain the temporal trajectory features of the vehicle and each intelligent agent around it; The interaction enhancement module is used to process temporal trajectory features to obtain interaction enhancement features; Furthermore, a gating calibration factor based on the relative motion state between the vehicle and multiple intelligent agents around it is introduced to dynamically reweight the initial weights of the standard multi-head self-attention calculation. The weighted aggregated features are then passed through a feedforward network to output interactive enhancement features. The multimodal prediction module is used to input interactive enhancement features into the multimodal prediction head for processing, generating multiple modal predictions of future trajectories and corresponding explicit probability distributions. The multimodal prediction head is built on a conditional variational autoencoder architecture and adds a probability estimation branch that runs in parallel with the decoder. The explicit probability value of each generated future trajectory is directly calculated through the probability estimation branch. The prediction output interface module is used to package multiple sets of formatted future trajectories and their explicit probability values, and send them to the vehicle's decision-making and path planning module.

Images