A multi-agent trajectory prediction method based on spatio-temporal causal interaction modeling

Abstract

The application provides a multi-agent trajectory prediction method based on space-time causal interaction modeling, and belongs to the behavior prediction field of multi-agent systems. The method solves the problem of low trajectory prediction accuracy caused by data bias and false association in existing methods. The method comprises the following steps: extracting two-dimensional position coordinates of a lane center line corresponding to vehicle driving; calculating the deviation degree of the vehicle position and the corresponding lane center line, modeling the spatial dependence relationship between vehicles, outputting the spatial interaction features of the vehicles and performing time series modeling to extract the time series features of the vehicles; feature encoding to output lane center line features; fusing the time series features and the lane center line features to generate vehicle fusion features; constructing a causal discovery network to output a causal probability matrix; based on the vehicle fusion features and the causal probability matrix, extracting the space-time causal interaction features between the vehicles and inputting the trajectory decoder to generate multiple possible future trajectory sequences of each vehicle. The method is used in the fields of artificial intelligence, robots and intelligent driving.

Description

Technical Field

[0001] This invention belongs to the field of behavior prediction technology for multi-agent systems, and in particular relates to a multi-agent trajectory prediction method based on spatiotemporal causal interaction modeling. Background Technology

[0002] Multi-agent trajectory prediction is a key technology in fields such as artificial intelligence, robotics, and autonomous driving. It provides core support for system safety, collaboration, and decision-making capabilities by accurately inferring the future motion states of multiple interacting objects. Taking autonomous driving as a typical application scenario as an example, the vehicle trajectory prediction module analyzes the future driving paths of surrounding traffic participants, directly impacting the reliability of risk identification and collision prevention. However, real-world traffic scenarios are highly dynamic and complex, with variable individual vehicle behaviors, dense environmental interactions, and factors such as sensor measurement noise and changing road topology, posing numerous challenges to trajectory prediction.

[0003] Existing vehicle trajectory prediction methods are mainly divided into three categories: physics-based methods, maneuver-based methods, and interaction-based methods. Physics-based methods rely on vehicle kinematic models (such as constant speed or acceleration models) to derive future trajectories from physical equations. They are computationally efficient in simple scenarios and short-term predictions, but struggle to capture complex interactive behaviors. For example, in multi-vehicle parallel scenarios at urban intersections, this method cannot reflect the dynamic adjustments made by vehicles due to avoidance or overtaking. Maneuver-based methods predict trajectories by recognizing driving behaviors (such as lane changing and following other vehicles). Their advantage lies in incorporating driving intentions, but they rely on accurate maneuver classification and are prone to misjudgment in sudden situations (such as emergency braking). Interaction-based methods consider the interdependence among all traffic participants in the scenario, such as modeling spatial relationships between vehicles using graph neural networks or social pooling models. They are more suitable for real-world complex scenarios, but these methods still have significant limitations.

[0004] The primary drawback of existing trajectory prediction methods lies in the lack of an effective mechanism for handling data bias. Historical vehicle trajectory data often contains measurement biases due to sensor noise, environmental occlusion, or positioning system errors; for example, GPS signal drift can cause a vehicle's position to deviate from the actual lane centerline. This bias accumulates in the feature extraction process, misleading the model to learn incorrect motion patterns. Taking interaction-based methods as an example, they directly use raw trajectory data to model vehicle relationships. If a vehicle's trajectory deviates abnormally due to noise, the model may incorrectly strengthen the interaction weights between that vehicle and other objects, thereby reducing prediction accuracy. Although some studies have attempted to smooth the data using filtering algorithms, they have failed to incorporate targeted corrections based on environmental constraints (such as lane geometry information), resulting in insufficient adaptability in dynamic scenarios.

[0005] Another key drawback of existing methods is the difficulty in distinguishing between genuine causal influences and spurious correlations. Interaction-based methods often rely on the statistical properties of trajectory data (such as spatial proximity or kinematic correlation) to model vehicle relationships, but statistical association is not equivalent to causal dependence. For example, in congested scenarios, two vehicles may exhibit kinematic correlation due to simultaneous deceleration, but there may be no direct causal influence (e.g., only controlled by the same traffic light). Such spurious associations can lead to model interference from irrelevant participants, reducing predictive interpretability. Specifically, existing methods, such as attention-based models, cannot explicitly identify causal chains between vehicles, leading to an overemphasis on weakly correlated objects in complex interactions while ignoring vehicles that truly exert critical influence (e.g., the causal effect of an emergency braking vehicle on a following vehicle). This deficiency not only limits predictive accuracy but also increases the risk of misjudgment in critical safety scenarios.

[0006] In summary, the shortcomings of existing trajectory prediction methods in data bias correction and causal interaction modeling severely limit their application effectiveness in high-risk scenarios such as autonomous driving. Data bias interferes with the reliability of feature extraction, while spurious associations weaken the specificity of interaction modeling; both factors contribute to low trajectory prediction accuracy and poor robustness. Summary of the Invention

[0007] In view of this, the present invention aims to propose a multi-agent trajectory prediction method based on spatiotemporal causal interaction modeling, so as to solve the problem of low trajectory prediction accuracy caused by data bias and false correlation in existing methods.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A multi-agent trajectory prediction method based on spatiotemporal causal interaction modeling, the method comprising: Step S1: Obtain historical observation trajectory data of all vehicles in the target traffic scene and map data in the traffic scene, and extract the two-dimensional position coordinate information of the lane centerline corresponding to the driving of each vehicle; Step S2: Calculate the deviation of the vehicle position from the corresponding lane centerline at each time point based on the two-dimensional position coordinate information of the lane centerline corresponding to each vehicle's movement. When the relative distance exceeds a preset threshold, a penalty weight is applied, and a self-attention mechanism is used to model the spatial dependency relationship between vehicles, outputting vehicle spatial interaction features based on lane centerline constraints. Step S3: Use a temporal neural network to perform temporal modeling of vehicle spatial interaction features and extract high-dimensional temporal features for each vehicle; Step S4: Perform feature encoding on the two-dimensional position coordinate sequence of the lane centerline and output the lane centerline features; Step S5: Fuse the high-dimensional temporal features of the vehicle with the lane centerline features to generate vehicle fusion features that include scene context information; Step S6: Construct a causal discovery network based on vehicle fusion features, distinguish between real causal influences and false spatial associations, and output the causal probability matrix between vehicles; Step S7: Based on vehicle fusion features and causal probability matrix, update features through a self-attention mechanism to extract spatiotemporal causal interaction features between vehicles; Step S8: Input the spatiotemporal causal interaction features between vehicles into the trajectory decoder. The decoder generates multiple possible future trajectory sequences for each vehicle through multi-step prediction.

[0009] Furthermore, a preferred embodiment is proposed, wherein step S1 includes: Obtain historical trajectory data of all vehicles participating in traffic on the road. ,in, Representative vehicle i At any moment t The status, including the vehicle's lateral and longitudinal position information at each moment. and lateral and longitudinal velocity information , N Indicates the number of vehicles. Indicates the historical observation time domain; Acquire the position data of the center lines of the surrounding lanes during vehicle operation. ,in, Representative vehicle i The corresponding lane centerline information, the lane centerline is composed of It consists of path points, and These represent the horizontal and vertical coordinates of the path points, respectively.

[0010] Furthermore, a preferred embodiment is proposed, wherein step S2 includes: Calculate vehicles i Position at time t closest point to the lane centerline relative distance d :

[0011] Set deviation thresholds based on road width and standard lane width. ; Determine the relative distance between the vehicle's position and the nearest point on the lane centerline. Does it exceed the set deviation threshold? If it exceeds, then the trajectory position data is determined. There is a deviation, and a position mask matrix is ​​constructed. The mask value is assigned according to the following rules: when hour, The corresponding weight is set to 1, otherwise it is set to 0. ; A self-attention mechanism is used to spatially model vehicle trajectory features, outputting vehicle spatial interaction features:

[0012] in, This indicates that elements at corresponding positions are multiplied. This represents the dimensions of the query matrix Q, the key matrix K, and the value matrix V.

[0013] Furthermore, a preferred method is proposed, wherein step S3 employs an LSTM network for timing coding, and the coding process is as follows:

[0014] in, This represents the temporal characteristics of all vehicles involved in traffic. This is an LSTM network used to encode the temporal features of vehicles.

[0015] Furthermore, a preferred method is proposed, wherein step S4 uses an LSTM network to encode the lane centerline position sequence, including:

[0016] in, For vehicles i The positional characteristics of the corresponding lane centerline. For an LSTM network used to encode lane centerline location features, For vehicles i Distribution of path points along the center lines of the surrounding lanes.

[0017] Furthermore, a preferred method is proposed, wherein the feature fusion in step S5 employs an aggregation operation to generate vehicle fused features:

[0018] in, For aggregation operations, For the acquired vehicle i The temporal characteristics; Integrating the fusion features of all traffic-participating vehicles yields .

[0019] Furthermore, a preferred approach is proposed, in which the causal discovery network in step S6 is implemented using an MLP to calculate the vehicle... i and vehicles j The probability of causal influence between them is:

[0020] in, It is a causal probability matrix. For vehicles j Vehicle fusion features that include scene context information.

[0021] Furthermore, a preferred approach is proposed, wherein the self-attention mechanism modeling process in step S7 is as follows:

[0022] in, For the spatiotemporal causal interaction characteristics of all traffic-participating vehicles, , and Let be a linear transformation matrix.

[0023] Based on the same inventive concept, the present invention also proposes a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes a multi-agent trajectory prediction method based on spatiotemporal causal interaction modeling as described in any of the preceding claims.

[0024] Based on the same inventive concept, the present invention also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of a multi-agent trajectory prediction method based on spatiotemporal causal interaction modeling as described above.

[0025] Compared with the prior art, the beneficial effects of the present invention are: First, the method proposed in this invention significantly improves the accuracy and reliability of trajectory prediction. In complex dynamic scenarios, such as urban road interactions, traditional methods suffer from deviations in historical trajectory data due to sensor noise or positioning errors, thus affecting the accuracy of feature extraction. This invention uses lane centerline geometric constraints to detect and weight trajectory deviations in real time, effectively suppressing the interference of abnormal data on model learning and making vehicle motion features more closely resemble real-world behavior. For example, when a vehicle's position deviates from its lane due to GPS drift, the model applies penalty weights through a mask matrix to reduce the impact of deviation data in spatial interaction modeling, thereby generating a predicted trajectory that better fits physical constraints. Experimental verification shows that on the ArgoverseMotion Forecasting dataset, the optimal predicted trajectory output by this invention significantly reduces the deviation from the real trajectory, especially demonstrating robust performance in challenging scenarios such as sharp turns and congested lane changes.

[0026] Secondly, the method proposed in this invention enhances the model's adaptability and robustness in real-world environments. Most existing methods rely on statistical association modeling of interactions, making them susceptible to spurious correlations, such as accidental synchronization of movements between unrelated vehicles. This invention integrates a causal discovery mechanism, which can explicitly distinguish between genuine causal dependencies between vehicles (such as the direct impact of a vehicle's braking on a following vehicle) and spurious spatial associations (such as parallel vehicles controlled by the same traffic light but without direct interaction), allowing the model to focus on key causal objects and avoid overfitting to irrelevant noise. This mechanism improves the model's ability to analyze complex interaction logic, ensuring that predicted trajectories in dynamic scenarios (such as intersection merging and emergency avoidance) both comply with traffic rules and preserve the diversity of individual intentions, thereby supporting safer decisions by autonomous driving systems.

[0027] Furthermore, the method proposed in this invention achieves efficient fusion and collaborative utilization of multi-source information. By organically integrating historical vehicle motion states, inter-vehicle interactions, and structured lane environment information through a hierarchical architecture, the model can capture deep dependencies in the spatiotemporal dimensions. Specifically, the lane centerline-vehicle interaction module embeds road topology constraints into vehicle features to generate a fusion representation rich in scene context; while the vehicle spatial interaction modeling module constrains attention weights through a causal probability matrix, enhancing the modeling efficiency of real interactions. This collaborative design avoids the information silo problem in traditional methods, ensuring that the prediction results simultaneously satisfy motion smoothness, environmental compliance, and interaction rationality. Attached Figure Description

[0028] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is an overall flowchart of a multi-agent trajectory prediction method based on spatiotemporal causal interaction modeling as described in this invention; Figure 2 This is a schematic diagram of the overall model structure described in this invention; Figure 3 This is a diagram illustrating the trajectory prediction effect described in this invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other, and the described embodiments are only some embodiments of the present invention, not all embodiments.

[0030] Implementation Method 1: This implementation method addresses the problem of low trajectory prediction accuracy caused by data bias and spurious correlations in existing trajectory prediction methods. It proposes a multi-agent trajectory prediction method based on spatiotemporal causal interaction modeling. The method includes: Step S1: Obtain historical observation trajectory data of all vehicles in the target traffic scene and map data in the traffic scene, and extract the two-dimensional position coordinate information of the lane centerline corresponding to the driving of each vehicle; Step S2: Calculate the deviation of the vehicle position from the corresponding lane centerline at each time point based on the two-dimensional position coordinate information of the lane centerline corresponding to each vehicle's movement. When the relative distance exceeds a preset threshold, a penalty weight is applied, and a self-attention mechanism is used to model the spatial dependency relationship between vehicles, outputting vehicle spatial interaction features based on lane centerline constraints. Step S3: Use a temporal neural network to perform temporal modeling of vehicle spatial interaction features and extract high-dimensional temporal features for each vehicle; Step S4: Perform feature encoding on the two-dimensional position coordinate sequence of the lane centerline and output the lane centerline features; Step S5: Fuse the high-dimensional temporal features of the vehicle with the lane centerline features to generate vehicle fusion features that include scene context information; Step S6: Construct a causal discovery network based on vehicle fusion features, distinguish between real causal influences and false spatial associations, and output the causal probability matrix between vehicles; Step S7: Based on vehicle fusion features and causal probability matrix, update features through a self-attention mechanism to extract spatiotemporal causal interaction features between vehicles; Step S8: Input the spatiotemporal causal interaction features between vehicles into the trajectory decoder. The decoder generates multiple possible future trajectory sequences for each vehicle through multi-step prediction.

[0031] Existing methods (such as filter-based smoothing or statistical modeling) typically use raw trajectory data directly, lacking targeted deviation correction mechanisms. The method proposed in this implementation utilizes prior environmental knowledge of the lane centerline to construct a dynamic deviation detection and correction process. By calculating the Euclidean distance between the vehicle position and the lane centerline and setting an adaptive threshold based on road geometry, a mask penalty is applied to trajectory data exhibiting deviations. This principle uses environmental constraints as a correction benchmark, suppressing the propagation of measurement errors at the source, rather than relying solely on data-driven post-processing. For example, traditional attention mechanisms might treat all trajectory points equally, while this implementation differentiates weights through a mask matrix, allowing the model to naturally weaken the influence of outliers during spatial interaction modeling, thereby improving the purity of feature extraction.

[0032] Existing methods (such as neural networks or social pooling) mostly model vehicle relationships based on spatial proximity or motion correlation, failing to distinguish between correlation and causation. The method proposed in this implementation introduces a causal discovery network (such as an MLP architecture) that explicitly quantifies the causal confidence of vehicle i's influence on vehicle j by learning the causal probability matrix between vehicle fusion features. This principle overcomes the limitations of statistical association, achieving end-to-end causal inference through differentiable networks. For example, in self-attention mechanisms, traditional methods rely solely on feature similarity to calculate weights, while this invention uses the causal probability matrix as an attention mask, forcing the model to focus on genuine causal chains, such as the direct influence of the preceding vehicle on the following vehicle in a car-following scenario, and suppressing spurious associations, such as unintentional motion synchronization of lateral vehicles. This causally constrained attention mechanism ensures the relevance and interpretability of interaction modeling.

[0033] Furthermore, the method proposed in this embodiment achieves progressive modeling from local features to global inference through the serial coupling of spatiotemporal causal interactions. First, individual trends are extracted through spatial interaction with bias correction and temporal encoding. Then, rich features are generated by fusing lane context. Finally, the essence of the interaction is extracted through causal discovery. This hierarchical refinement principle avoids the feature confusion problem common in existing methods, namely, avoiding the misjudgment of environmental noise as interaction signals, enabling the model to focus on key information layer by layer.

[0034] Implementation Method 2, see below Figures 1 to 3 This embodiment describes a complete implementation process for a multi-agent trajectory prediction method based on spatiotemporal causal interaction modeling as described in Embodiment 1, including: Step 1: Data Acquisition Steps: Acquire historical observation trajectory data of all vehicles in the target traffic scene, including the lateral and longitudinal position coordinates and lateral and longitudinal speeds of each vehicle during the observation period; at the same time, acquire high-precision map data of the target traffic scene and extract the two-dimensional position coordinate information of the lane centerline corresponding to the driving of each vehicle.

[0035] In this embodiment, step 1 includes the following steps: Step 1.1: Obtain historical trajectory data of all vehicles participating in traffic on the road. ,in Representative vehicle i At any moment t The status, including the vehicle's lateral and longitudinal position information at each moment. and lateral and longitudinal velocity information , N Indicates the number of vehicles. This indicates the historical observation time domain.

[0036] Step 1.2: Obtain lane centerline position data around the vehicle during its movement. ,in, Representative vehicle i The corresponding lane centerline information, the lane centerline is composed of It consists of path points, and These represent the horizontal and vertical coordinates of the path points, respectively.

[0037] Step 2, Vehicle Trajectory Deviation Detection and Spatial Interaction Modeling: Based on the data from Step 1, calculate the degree of deviation between the vehicle's position and the corresponding lane centerline at each time point. A deviation threshold is preset based on the traffic rules and road geometry of the target scenario. When the deviation exceeds the threshold at a certain time, a penalty weight is applied to the trajectory data at that time to suppress the negative impact of data deviation on feature extraction. The penalty weight for the deviated trajectory is input into the vehicle spatial interaction modeling module to obtain vehicle spatial features based on lane centerline constraints.

[0038] In this embodiment, step 2 includes the following steps: Step 2.1: At each historical observation time, calculate the perpendicular distance from the vehicle's position to the center line of the corresponding lane. The projection point of the vehicle's position onto the lane centerline is determined using vector projection. The Euclidean distance between the projection point and the vehicle's actual position is then calculated. Specifically, it includes the following sub-steps: Step 2.1.1: Based on Step 1, for each historical moment Observed vehicles i Location information is ,vehicle i The corresponding lane centerline information can be represented as a series of points. .

[0039] Step 2.1.2: For the vehicle i Location Find the nearest point on the center line of the lane. Calculate the distance between two locations. :

[0040] Step 2.2: Set the deviation threshold The threshold Determined based on engineering experience regarding lane width and vehicle width. If the vertical distance... Exceeding the deviation threshold Determine the vehicle trajectory location data at that moment. There is a discrepancy.

[0041] Step 2.3: Construct the position mask matrix The matrix elements characterize the reliability of vehicle historical trajectory data. The value is determined by the distance from the vehicle's position to the center line of the lane. With threshold The deviation between them is used to determine this. Specifically, when At that time, the position mask matrix The corresponding mask value is set to 1 when At that time, the position mask matrix The size of the corresponding mask value is set to Add penalty constraints to the deviation trajectory data.

[0042] Step 2.4: Encode the trajectory features of the N vehicles observed at each historical moment into a feature matrix Xt=sit|i=1, A self-attention mechanism is employed to encode the spatial features of vehicles at each historical observation time to obtain the updated spatial interaction features of all participating vehicles. The vehicle trajectory feature matrix is ​​obtained through a linear transformation matrix. , and These are mapped to the query matrix Q, the key matrix K, and the value matrix V, respectively. The mapping relationships are as follows:

[0043] A self-attention mechanism is used to model the spatial dependencies between vehicles. The query matrix Q and the key matrix K are multiplied by a dot product, and the attention weights are normalized using the Softmax function. Then, a location mask matrix is ​​introduced. After weighted correction, the spatial characteristics of all vehicles involved in the traffic are finally obtained. for:

[0044] in, This indicates that elements at corresponding positions are multiplied. The dimensions of the query matrix Q, key matrix K, and value matrix V can be set to 1. .

[0045] Step 3, Vehicle Temporal Feature Modeling: Since LSTM networks have better prediction performance and computational efficiency when processing short-term time series data, based on the vehicle spatial interaction features obtained in Step 2.2... A single-layer LSTM network is used for temporal encoding to capture the motion trend, ultimately outputting high-dimensional temporal features for each vehicle. The encoding process is as follows:

[0046] in, This represents the temporal characteristics of all vehicles involved in traffic. This is an LSTM network used to encode the temporal features of vehicles.

[0047] Step 4: Lane Centerline Feature Extraction: Perform temporal modeling of the lane centerline position information for each vehicle during its journey, and extract the lane centerline position features. For each vehicle... i Distribution of path points on the center line of the surrounding lanes As input, a single-layer LSTM network is used for feature extraction. The calculation process is as follows:

[0048] in, For vehicles i The positional characteristics of the corresponding lane centerline. This is an LSTM network used to encode lane centerline position features. Further, the lane centerline position feature matrix corresponding to all participating vehicles can be obtained as follows: .

[0049] Step 5, Lane Centerline-Vehicle Interaction Modeling Step: For the vehicle obtained in Step 3... i Temporal characteristics The corresponding lane centerline features obtained in step 4 Perform feature fusion to obtain vehicle fused features that include scene context information. :

[0050] in, This is an aggregation operation. Further, the fusion features of all participating vehicles are integrated to obtain... .

[0051] Step 6: Causal Discovery Between Vehicles: An MLP network is used to construct a causal discovery model. The vehicle fusion features output from Step 5 are input into the MLP network to obtain the vehicle... i and vehicles j The probability of causal influence between them is:

[0052] in, This is a causal probability matrix, where the matrix elements represent vehicles. i For vehicles j The confidence level for the existence of a causal relationship ranges from [0,1]. For vehicles j Vehicle fusion features that include scene context information.

[0053] Step 7, Vehicle-to-Vehicle Interaction Modeling Step: Based on the vehicle features output in Step 5 and the causal probability matrix between vehicles obtained in step 6 A self-attention mechanism model is performed, and further interaction modeling is conducted on the spatial distribution relationships between vehicles to extract spatiotemporal causal interaction features between vehicles. The causal probability matrix is ​​used as a constraint term for the attention weights to perform a weighted update of the vehicle feature matrix. The calculation process is as follows:

[0054] in, This is a spatiotemporal causal interaction feature for all vehicles involved in traffic, which includes vehicle temporal trends, lane scene constraints, and information on the actual causal influence between vehicles.

[0055] Step 8: Trajectory Prediction Decoding Output: Using an LSTM network as the decoder, the spatiotemporal causal interaction features of all participating vehicles obtained in Step 7 are processed. The input is fed into an LSTM decoder, which outputs multiple possible vehicle trajectory information in the future time domain. The decoding process is as follows:

[0056] in, For vehicles i In the future t Predicted location at time Represents the predicted duration. This is an LSTM network used for decoding vehicle predicted trajectories.

[0057] To verify the performance of the multi-agent trajectory prediction method based on spatiotemporal causal interaction modeling proposed in this invention, this section uses urban road scene data from the Argoverse Motion Forecasting dataset for experimental verification, with the parameters set as follows: observation duration 2s and prediction duration 3s. Figure 3 The model's predictive performance in typical urban road scenarios is demonstrated. The visualization rules are as follows: solid black lines represent the vehicle's historical observed trajectory, dashed red lines represent the vehicle's actual future trajectory, and green lines represent multiple predicted trajectories output by the model. The green lines marked with a blue star represent the most probable predicted trajectory. It can be seen that the model of this invention can output multiple future trajectories that conform to traffic rules and motion logic, and the optimal predicted trajectory marked with a star deviates little from the actual trajectory, fully demonstrating its effectiveness in generating reasonable predicted trajectories in complex real-world traffic scenarios.

[0058] This invention proposes a multi-agent trajectory prediction method based on spatiotemporal causal interaction modeling. By collaboratively utilizing historical motion states, inter-agent interactions, and structured environmental information, and introducing an environmental constraint-based deviation correction mechanism and a causal discovery-based interaction modeling mechanism, it achieves high-precision and robust trajectory prediction. On one hand, this method uses environmental geometric constraints to accurately identify and weightedly correct historical trajectory deviations, effectively suppressing the interference of trajectory anomalies caused by sensor measurement noise on feature extraction, and improving the model's adaptability in real and complex perception environments. On the other hand, by introducing a causal discovery mechanism, it can distinguish between genuine causal influences and spurious spatial correlations between agents, enabling the model to focus on a few agents exerting key influences and avoiding interference from a large number of irrelevant or weakly correlated interaction objects, thus strengthening the effectiveness of inter-agent interaction modeling. In summary, this invention effectively improves the problem of insufficient trajectory prediction accuracy caused by trajectory data deviations and spurious correlations. In the typical application scenario of intelligent driving, this method can provide reliable technical support for vehicle trajectory prediction and autonomous driving decision-making and planning. Furthermore, its core framework also has the potential to be migrated and applied to other multi-agent systems such as robot collaboration.

[0059] Implementation Method 3: This implementation method proposes a computer device, including a memory and a processor. The memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes a multi-agent trajectory prediction method based on spatiotemporal causal interaction modeling as described in any one of Implementation Methods 1 to 2.

[0060] Implementation Method 3: This implementation method proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of a multi-agent trajectory prediction method based on spatiotemporal causal interaction modeling as described in any one of Implementation Methods 1 to 2.

[0061] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0062] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0063] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this disclosure and not to limit its protection scope. Although this disclosure has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading this disclosure, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the protection scope of the published pending claims.

Claims

1. A multi-agent trajectory prediction method based on spatiotemporal causal interaction modeling, characterized in that, The method includes: Step S1: Obtain historical observation trajectory data of all vehicles in the target traffic scene and map data in the traffic scene, and extract the two-dimensional position coordinate information of the lane centerline corresponding to the driving of each vehicle; Step S2: Calculate the deviation of the vehicle position from the corresponding lane centerline at each time point based on the two-dimensional position coordinate information of the lane centerline corresponding to each vehicle's movement. When the relative distance exceeds a preset threshold, a penalty weight is applied, and a self-attention mechanism is used to model the spatial dependency relationship between vehicles, outputting vehicle spatial interaction features based on lane centerline constraints. Step S3: Use a temporal neural network to perform temporal modeling of vehicle spatial interaction features and extract high-dimensional temporal features for each vehicle; Step S4: Perform feature encoding on the two-dimensional position coordinate sequence of the lane centerline and output the lane centerline features; Step S5: Fuse the high-dimensional temporal features of the vehicle with the lane centerline features to generate vehicle fusion features that include scene context information; Step S6: Construct a causal discovery network based on vehicle fusion features, distinguish between real causal influences and false spatial associations, and output the causal probability matrix between vehicles; Step S7: Based on vehicle fusion features and causal probability matrix, update features through a self-attention mechanism to extract spatiotemporal causal interaction features between vehicles; Step S8: Input the spatiotemporal causal interaction features between vehicles into the trajectory decoder. The decoder generates multiple possible future trajectory sequences for each vehicle through multi-step prediction. Step S2 includes: Calculate vehicles i Position at time t closest point to the lane centerline relative distance d : Set deviation thresholds based on road width and standard lane width. ; Determine the relative distance between the vehicle's position and the nearest point on the lane centerline. Does it exceed the set deviation threshold? If it exceeds, then the trajectory position data is determined. There is a deviation, and a position mask matrix is ​​constructed. The mask value is assigned according to the following rules: when hour, The corresponding weight is set to 1, otherwise it is set to 0. ; A self-attention mechanism is used to spatially model vehicle trajectory features, outputting vehicle spatial interaction features: in, Multiply the elements at corresponding positions. To query the dimensions of matrix Q, key matrix K, and value matrix V.

2. The multi-agent trajectory prediction method based on spatiotemporal causal interaction modeling according to claim 1, characterized in that, Step S1 includes: Obtain historical trajectory data of all vehicles participating in traffic on the road. ,in, Representative vehicle i At any moment t The status, including the vehicle's lateral and longitudinal position information at each moment. and lateral and longitudinal velocity information , N Indicates the number of vehicles. Indicates the historical observation time domain; Acquire the position data of the center lines of the surrounding lanes during vehicle operation. ,in, Representative vehicle The corresponding lane centerline information.

3. The multi-agent trajectory prediction method based on spatiotemporal causal interaction modeling according to claim 1, characterized in that, Step S3 uses an LSTM network for timing coding. The coding process is as follows: in, For the temporal characteristics of all vehicles involved in traffic, This is an LSTM network used to encode the temporal features of vehicles.

4. The multi-agent trajectory prediction method based on spatiotemporal causal interaction modeling according to claim 1, characterized in that, Step S4 uses an LSTM network to encode the lane centerline position sequence, including: in, For vehicles i The corresponding positional characteristics of the lane centerline, For an LSTM network used to encode lane centerline location features, For vehicles i Distribution of path points along the center lines of the surrounding lanes.

5. The multi-agent trajectory prediction method based on spatiotemporal causal interaction modeling according to claim 4, characterized in that, The feature fusion in step S5 employs an aggregation operation to generate vehicle fused features: in, For aggregation operations, For the acquired vehicle i The temporal characteristics; Integrating the fusion features of all traffic-participating vehicles yields .

6. The multi-agent trajectory prediction method based on spatiotemporal causal interaction modeling according to claim 5, characterized in that, The causal discovery network in step S6 is implemented using MLP to calculate the vehicle... i and vehicles j The probability of causal influence between them is: in, It is a causal probability matrix. For vehicles j Vehicle fusion features that include scene context information.

7. The multi-agent trajectory prediction method based on spatiotemporal causal interaction modeling according to claim 6, characterized in that, The self-attention mechanism modeling process in step S7 is as follows: in, For the spatiotemporal causal interaction characteristics of all traffic-participating vehicles, , and Let be a linear transformation matrix.

8. A computer device, characterized in that: The system includes a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes a multi-agent trajectory prediction method based on spatiotemporal causal interaction modeling as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of a multi-agent trajectory prediction method based on spatiotemporal causal interaction modeling as described in any one of claims 1-7.

Images