A vehicle joint trajectory prediction method and system

Abstract

The application provides a vehicle joint trajectory prediction method and system, and relates to the technical field of automatic driving vehicle trajectory prediction. The method and system improve the prediction accuracy, physical consistency and robustness in a complex traffic environment by eliminating environmental mixed factor interference through causal semantic interaction modeling, constructing a dynamic risk potential field and performing iterative optimization, and explicitly mine the causal dependence relationship of vehicle behavior and integrate the physical safety constraints of risk perception.

Description

Technical Field

[0001] This application relates to autonomous vehicle trajectory prediction technology, and in particular to a vehicle joint trajectory prediction method and system. Background Technology

[0002] Accurate vehicle trajectory prediction is a core element in ensuring the safe operation of autonomous driving systems and enhancing the passenger experience. In scenarios with sparse traffic participants and simple road structures, existing prediction algorithms can provide relatively reliable trajectory estimates.

[0003] However, when faced with complex traffic environments involving high-density vehicle interactions, such as dense urban traffic flow or multi-vehicle parallel sections of highways, the prediction accuracy of current methods deteriorates sharply, posing a serious risk to autonomous driving decision-making systems. This performance degradation stems from the high interdependence of traffic participant behaviors: the future trajectory of an individual vehicle is determined not only by its own historical motion state but also by the combined influence of changes in the intentions of surrounding vehicles, road topology, and environmental dynamics. Early prediction methods based on physical models constructed prediction maps through explicit kinematic equations. While these methods could enforce basic physical constraints and achieve efficient inference, their implicit assumptions of low-order motion continuity and weak interaction coupling simplifications meant that they were only suitable for short-term predictions and simple scenarios, failing to capture nonlinear behavioral patterns in complex interactions. To overcome this limitation, researchers turned to data-driven joint prediction methods, which model the entire traffic scenario as a unified system and generate spatiotemporally consistent collaborative trajectories by analyzing the game relationships between vehicles. Although joint prediction models based on recurrent neural networks, Transformers, or graph neural networks perform well in conventional scenarios, they are essentially black-box architectures that fit statistical dependencies in data, exhibiting two key drawbacks. First, these models lack genuine causal reasoning capabilities, relying excessively on superficial statistical correlations in the training data, leading to sensitivity to environmental confounding factors. Second, existing methods lack physical reasoning and safety constraint mechanisms, focusing solely on minimizing displacement errors while neglecting the physical feasibility and safety of the trajectory. As a result, the generated predicted trajectories may violate vehicle kinematic limits, such as sharp turns exceeding tire grip limits, or overlapping with other vehicles, creating potential collision risks. For safety-critical autonomous driving systems, such risk-aware predictions are completely unacceptable. Furthermore, other limitations exist in implementation: fully connected interaction graphs or global attention mechanisms are prone to overfitting to dominant interaction patterns; objective-conditional methods rely on the premise of identifiable intent, failing in highly ambiguous interaction scenarios; and methods based on partially observable Markov games are limited by computational complexity and the difficulty of designing reward functions. Therefore, there is an urgent need for a unified framework that can explicitly mine causal dependencies in vehicle behavior from first principles and integrate risk-aware physical safety constraints into the prediction process to address the fundamental problems of insufficient prediction accuracy, physical consistency, and robustness in complex driving environments.

[0004] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0005] This application provides a vehicle joint trajectory prediction method and system, which has the advantages of improving prediction accuracy, physical consistency and robustness in complex traffic environments, while explicitly mining the causal dependencies of vehicle behavior and integrating physical safety constraints of risk perception.

[0006] Firstly, the vehicle joint trajectory prediction method provided in this application adopts the following technical solution: A vehicle joint trajectory prediction method includes: Obtain raw traffic scene data, perform spatiotemporal feature decoupling encoding on the data, and obtain a contextual representation that integrates vehicle dynamic behavior and map static topology; Based on the aforementioned context representation, a vehicle semantic topology map is constructed in the Frenet coordinate system, and causal semantic interaction modeling is completed through a backdoor adjustment mechanism to eliminate interference from environmental confounding factors and obtain causal attention weights. Based on the causal attention weights, a joint decoder is used to generate a multimodal initial joint trajectory, and a dynamic risk potential field is constructed based on the predicted positions of semantic neighbor vehicles. Using the dynamic risk potential field as an energy function, and starting from the initial multimodal joint trajectory, the vehicle joint trajectory is output through iterative optimization along the risk gradient direction using Langevin dynamics.

[0007] Optionally, the spatiotemporal feature decoupling encoding specifically involves: embedding Fourier features into the vehicle's continuous kinematic state vector, independently extracting the vehicle's historical trajectory temporal features and vectorized map topological features through a two-stream architecture, and then fusing them through a bidirectional cross-attention mechanism to obtain the context representation.

[0008] Optionally, Fourier feature embedding involves concatenating the kinematic state vector composed of vehicle position, velocity, and heading angle through cosine and sine function mappings, and then mapping it to a high-dimensional embedding space through a learnable linear projection matrix.

[0009] Optionally, the vehicle semantic topology graph uses a geometric relationship classifier to determine the front, rear, left, and right semantic spatial relationships between the target vehicle and its neighboring vehicles, and constructs a directed semantic interaction graph based on these semantic spatial relationships.

[0010] Optionally, the causal semantic interaction modeling specifically involves estimating potential confounding variables in vehicle interaction through variational inference, removing confounding bias terms from the conventional attention score based on the backdoor adjustment criterion, and calculating the causal attention weights.

[0011] Optionally, the joint decoder is a joint GRU decoder, which generates scene-level multimodal initial joint trajectories synchronously based on the causal attention weights autoregressively.

[0012] Optionally, the dynamic risk potential field is formed by the superposition of Gaussian repulsive forces generated by surrounding semantic neighbor vehicles, and the collision risk is dynamically quantified based on the vehicle prediction position of the initial joint trajectory.

[0013] Optionally, Langevin dynamics iterative optimization is as follows: starting from the initial predicted trajectory point, update the trajectory point in the opposite direction of the dynamic risk potential energy gradient, and add random noise to complete trajectory fine-tuning.

[0014] Secondly, this application provides a vehicle joint trajectory prediction system, comprising: The data acquisition module is used to acquire raw traffic scene data, perform spatiotemporal feature decoupling encoding on the data, and obtain a contextual representation that integrates vehicle dynamic behavior and map static topology; The topology module is used to construct a vehicle semantic topology map in the Frenet coordinate system based on the context representation, and to complete causal semantic interaction modeling through a backdoor adjustment mechanism, thereby eliminating interference from environmental confounding factors to obtain causal attention weights. The dynamic risk potential field module is used to generate a multimodal initial joint trajectory based on the causal attention weights using a joint decoder, and to construct a dynamic risk potential field based on the predicted positions of semantic neighbor vehicles. The output module is used to take the dynamic risk potential field as an energy function, take the multimodal initial joint trajectory as the starting point, and iteratively optimize it along the risk gradient direction through Langevin dynamics to output the vehicle joint trajectory.

[0015] In summary, this application eliminates environmental confounding factors through causal semantic interaction modeling and constructs a dynamic risk potential field for iterative optimization, thereby improving prediction accuracy, physical consistency, and robustness in complex traffic environments. At the same time, it explicitly mines the causal dependencies of vehicle behavior and integrates the physical safety constraints of risk perception. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the first embodiment of the vehicle joint trajectory prediction method of this application; Figure 2 This is a structural block diagram of the first embodiment of the vehicle joint trajectory prediction system of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0018] Traditional vehicle trajectory prediction methods show a significant drop in accuracy in complex traffic scenarios with dense interactions, leading to a degradation in autonomous driving performance. Existing data-driven models lack true causal reasoning capabilities, rely excessively on statistical correlations, are susceptible to spurious correlations, and have poor generalization ability; at the same time, they lack physical reasoning and safety constraints, and are prone to generating physically infeasible trajectories.

[0019] To address this, this application provides a vehicle joint trajectory prediction method, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the vehicle joint trajectory prediction method of this application.

[0020] In this embodiment, the vehicle joint trajectory prediction method includes the following steps: Step S10: Obtain the original traffic scene data, perform spatiotemporal feature decoupling encoding on the data, and obtain a contextual representation that integrates vehicle dynamic behavior and map static topology.

[0021] To facilitate understanding of this embodiment, some key technical concepts involved are explained below: Raw traffic scene data refers to real-time information about the vehicle itself and its surrounding environment, obtained from onboard sensors or high-precision maps, such as the vehicle's position, speed, heading, lane information, and traffic signs. This data is the basic input for trajectory prediction.

[0022] Spatiotemporal feature decoupling coding is a data processing method that aims to independently encode the temporal series features in the historical trajectory of a vehicle and the spatial features in the map topology, in order to more accurately capture the relationship between vehicle behavior and the environment.

[0023] The context representation is a comprehensive feature vector that integrates vehicle dynamic behavior information and map static topology information after spatiotemporal feature decoupling encoding, which can comprehensively describe the current traffic scene.

[0024] The Frenet coordinate system is a local coordinate system commonly used in the field of autonomous driving. Its coordinate axes are usually aligned with the center line of the road, which helps to simplify the description of the vehicle's motion within the lane.

[0025] A vehicle semantic topology graph is a graph structure used to represent the semantic interactions between vehicles in a traffic scenario. Nodes in the graph represent vehicles, and edges represent the semantic relationships between them.

[0026] Backdoor adjustment mechanism is a causal inference method used to estimate the causal effect of treatment variables on outcome variables in the presence of confounding variables, thereby obtaining causal relationships.

[0027] Causal semantic interaction modeling is a process of using backdoor adjustment mechanisms to perform causal analysis on semantic interactions between vehicles, aiming to identify and quantify the real causal dependencies between vehicle behaviors.

[0028] Causal attention weights are the output of causal semantic interaction modeling, representing the causal influence of different neighboring vehicles on the future behavior of the target vehicle after removing environmental confounding factors.

[0029] A joint decoder is a neural network architecture used to generate future trajectories of multiple traffic participants based on the contextual representation of the input and causal attention weights.

[0030] The multimodal initial joint trajectory refers to a set of possible future joint trajectories generated by the joint decoder, reflecting the diversity of vehicle behavior.

[0031] Dynamic risk potential field is an energy field constructed in real time to quantify potential collision risks in traffic scenarios. Its energy value is higher in potential collision areas.

[0032] In optimization problems, the energy function is used to measure the "goodness" or "badness" of a state or trajectory. In this method, the dynamic risk potential field is used as the energy function.

[0033] Langevin dynamics is a set of stochastic differential equations used to simulate the motion of particles in a potential field. In this method, it is used to iteratively optimize the initial joint trajectory.

[0034] The risk gradient direction refers to the direction in which energy decreases most rapidly in the dynamic risk potential field. Updating the trajectory along this direction can gradually move the vehicle's trajectory away from the high-risk area.

[0035] The vehicle joint trajectory is the final predicted trajectory obtained after Langevin dynamics optimization, taking into account the causal interactions between vehicles and physical safety constraints.

[0036] It should be noted that the spatiotemporal feature decoupling encoding specifically involves: embedding Fourier features into the vehicle's continuous kinematic state vector, independently extracting the vehicle's historical trajectory temporal features and vectorized map topological features through a two-stream architecture, and then fusing them through a bidirectional cross-attention mechanism to obtain the context representation.

[0037] Vehicle continuous kinematic state vectors typically contain continuously changing physical quantities such as position, velocity, acceleration, and heading angle. Directly using these raw vectors as input may make it difficult for the model to capture their inherent periodicity, nonlinearity, and high-frequency information. Fourier feature embedding is a technique that maps low-dimensional continuous input data to a high-dimensional feature space. By introducing a combination of periodic functions (such as sine and cosine functions), it can effectively enhance the model's ability to perceive complex patterns in the input data. This embedding method allows the model to better understand the subtle changes in the vehicle's trajectory, providing richer representations for subsequent temporal feature extraction.

[0038] By employing the aforementioned technical solutions, Fourier feature embedding of the vehicle's continuous kinematic state vector effectively enhances the representation of vehicle dynamic behavior, enabling more precise capture of motion patterns in subsequent processing. Simultaneously, a dual-stream architecture is used to independently extract the vehicle's historical trajectory temporal features and vectorized map topological features, ensuring the purity of dynamic and static information in the initial stage and avoiding information confusion or loss. Building upon this, a bidirectional cross-attention mechanism is used for fusion, allowing for intelligent and adaptive interaction between the two modalities. Vehicle dynamic behavior can selectively utilize map topological information, while map topological information can also contribute to the understanding of vehicle motion intentions. This results in a more comprehensive and accurate fusion of vehicle dynamic behavior and map static topology, providing a high-quality and robust foundation for subsequent joint vehicle trajectory prediction, significantly improving prediction accuracy and adaptability to complex traffic scenarios.

[0039] It is understandable that Fourier feature embedding involves mapping the kinematic state vector of vehicle position, speed, and heading angle to a high-dimensional embedding space through a learnable linear projection matrix after concatenating the vectors using cosine and sine functions.

[0040] After mapping all kinematic state variables using cosine and sine functions and concatenating them into a composite feature vector, this composite feature vector is further transformed using a learnable linear projection matrix. This linear projection matrix maps the concatenated feature vector into a higher-dimensional embedding space. Through this linear transformation, the model can learn the weight relationships between different feature combinations, thereby finding a more discriminative representation for trajectory prediction in a higher-dimensional space. This high-dimensional embedding space provides richer and more abstract inputs for subsequent spatiotemporal feature decoupling encoding, helping to improve the model's understanding of complex traffic scenarios.

[0041] In practical implementation, in order to overcome the spectral bias of traditional multilayer perceptrons and improve the ability to capture high-frequency dynamic features in trajectory evolution, and to acquire target vehicle data, [further details are needed]. exist Continuous kinematic state vector at time step (Includes attributes such as position, velocity, and heading angle). Using a Fourier feature embedding function. Map it to a high-dimensional function space: in, Gaussian distribution from the control spectral range The random frequency matrix sampled in the middle; It is a learnable linear projection matrix used to map the concatenated high-dimensional features to... In the embedded space of dimensions.

[0042] Dual-stream independent coding of time-series and map information includes: (1) Time-series flow (vehicle history behavior coding): for vehicles Historical state embedding sequence A multi-head temporal self-attention mechanism is used to model its temporal evolution and output behavioral feature representations. : in, , , These are respectively query, key, and value vectors; It is a mask matrix used to block future information to ensure temporal causal consistency.

[0043] (2) Map flow (topological coding): The road network is represented as a set of polyline segments, and each polyline segment is discretized into a vector containing the coordinates of the start and end points and semantic attributes (such as lane type). A PointNet-based subgraph aggregation network is used to extract polyline-level features. As a map topology embedding.

[0044] To integrate vehicle motion intent with environmental topological constraints, in the context of vehicle behavior characteristics and map topological features A symmetrical bidirectional cross-attention mechanism is introduced between them to output the fused context representation. : in, This represents the cross-attention from the vehicle to the map. This indicates the reverse cross-attention from the map to the vehicle.

[0045] Step S20: Based on the context representation, construct a vehicle semantic topology map in the Frenet coordinate system, and complete causal semantic interaction modeling through the backdoor adjustment mechanism to remove environmental confounding factors and obtain causal attention weights; It is understandable that the vehicle semantic topology graph uses a geometric relationship classifier to determine the front, rear, left, and right semantic spatial relationships between the target vehicle and its neighboring vehicles, and constructs a directed semantic interaction graph based on these semantic spatial relationships.

[0046] Through the above technical solution, this embodiment can accurately determine the front, rear, left, and right semantic spatial relationships between the target vehicle and its neighboring vehicles, thereby transforming abstract inter-vehicle interactions into concrete and quantifiable semantic information. Based on these explicit semantic spatial relationships, a directed semantic interaction graph is constructed, clearly representing the mutual influence and potential interaction patterns between vehicles. This refined semantic topology graph construction method provides more accurate and richer input for subsequent causal semantic interaction modeling, effectively avoiding prediction biases caused by ambiguity or lack of interaction relationships, and significantly improving the accuracy and robustness of joint trajectory prediction.

[0047] In practice, existing data-driven methods often use Euclidean distance for mapping, which can easily learn spurious correlations caused by potential environmental confounding factors. This step aims to uncover the true causal relationships between vehicles.

[0048] The steps for constructing a semantic topology graph include: Using a geometric relation classifier in the Frenet (natural) coordinate system Identify the target vehicle with neighbor's vehicle Spatial semantic relationships between Construct a directed semantic interaction graph This mechanism introduces a strong inductive bias that aligns with real-world traffic logic (e.g., prioritizing the influence of leading vehicles ahead).

[0049] Variational inference of potential confounding factors includes: Due to unobserved variables (Such as weather, blind spot obstacles) can simultaneously affect neighboring vehicles. and target car Forming a backdoor path This embodiment utilizes variational inference to approximate the distribution of confounding variables, based on the global scene context. Predict its mean and variance, and sample to obtain potential confounding variables. .

[0050] Backdoor adjustments and causal attention calculations include: Based on Pearl's backdoor adjustment criterion, a causal graph attention layer is designed. The true causal intervention attention score is estimated by subtracting confounding bias terms. The first term is the standard attention score; the second term is the estimated confounding bias. The confounding suppression coefficient is denoted as . To estimate the confounding bias in the neural network, the scores are finally normalized using the Softmax function to obtain the final causal attention weights after eliminating spurious correlations. : Step S30: Based on the causal attention weights, a multimodal initial joint trajectory is generated using a joint decoder, and a dynamic risk potential field is constructed based on the predicted positions of semantic neighbor vehicles.

[0051] It is understandable that the causal semantic interaction modeling specifically involves estimating potential confounding variables in vehicle interaction through variational inference, removing confounding bias terms from the conventional attention score based on the backdoor adjustment criterion, and calculating the causal attention weights.

[0052] Specifically, variational inference, a powerful probabilistic inference tool, is used in this embodiment to approximate the posterior distribution of latent confounding variables that are difficult to observe directly during vehicle interaction. These latent confounding variables may encompass various factors that are not explicitly modeled but influence vehicle behavior, such as the driver's personalized driving style, road surface slippage, and sudden obstacles. By constructing a probabilistic model that includes these latent variables and iteratively optimizing it using variational inference algorithms, the system can effectively quantify the contribution of these latent confounding variables to vehicle interaction behavior, thereby identifying which seemingly related interactions are actually driven by common, hidden factors.

[0053] This embodiment introduces a backdoor adjustment criterion to accurately eliminate non-causal bias terms introduced by the aforementioned potential confounding variables in the conventional attention score. As a core method in causal inference, the backdoor adjustment criterion works by conditionalizing or adjusting specific confounding variables to isolate and identify the true causal effects between variables. In vehicle interaction modeling, conventional attention mechanisms may inadvertently capture spurious associations caused by confounding variables when calculating interaction scores between vehicles. By applying the backdoor adjustment criterion, the system can systematically eliminate the interference of these confounding variables on the attention score. For example, this can be achieved by introducing a compensation term to offset the influence of confounding variables when calculating attention weights, or by adjusting the attention score through a weighted average of different states of the confounding variables.

[0054] Through the above technical solution, this embodiment effectively addresses the attention bias problem caused by potential confounding variables in vehicle interaction modeling. Specifically, by estimating potential confounding variables in vehicle interaction through variational inference, the system can identify and quantify factors that are not directly observed but influence vehicle behavior. Based on this, using a backdoor adjustment criterion, the system can accurately remove non-causal bias terms caused by these confounding variables from the regular attention score. This ensures that the final calculated causal attention weights more realistically and accurately reflect the causal interaction relationships between vehicles, rather than merely superficial statistical correlations. Therefore, the initial joint trajectory generated based on these more reliable causal attention weights will provide more accurate and robust predictions of future vehicle behavior, significantly improving the accuracy and safety of trajectory prediction in complex traffic scenarios, especially in autonomous driving decisions requiring refined interaction understanding.

[0055] It should be noted that the joint decoder is a joint GRU decoder, which generates scene-level multimodal initial joint trajectories synchronously based on the causal attention weights autoregressive.

[0056] The joint GRU decoder is a recurrent neural network (RNN) architecture specifically designed for processing and generating sequential data. It effectively manages information flow through gating mechanisms such as update and reset gates, enabling it to capture long-distance temporal dependencies while maintaining computational efficiency. In this scheme, the decoder is configured to simultaneously handle trajectory generation tasks for multiple vehicles in a scene, taking into account their interdependencies to achieve "joint" trajectory prediction.

[0057] In practice, this step formulates the prediction problem as a joint distributed learning task and introduces physical collision constraints into the generated candidate trajectory space.

[0058] Joint multimodal trajectory generation includes: Using a joint GRU (Gated Cyclic Unit) decoder, scene-level multimodal joint trajectories are generated synchronously and autoregressively. For the first... The model outputs the corresponding mixture probability for each prediction pattern. and the initial predicted trajectory .

[0059] Dynamic risk potential field modeling includes: Purely data-driven initial trajectories may result in physical collisions. This embodiment introduces a physics-inspired dynamic risk potential field. The predicted location of the target vehicle Placed in a potential field, its potential energy is generated by the surrounding semantic neighbor vehicles (located at...). The potential field, formed by the superposition of Gaussian repulsive forces generated by the vehicle, establishes a high-energy barrier near surrounding vehicles, dynamically quantifying the potential collision risk. Step S40: Using the dynamic risk potential field as an energy function, and starting from the initial multimodal joint trajectory, iteratively optimize along the risk gradient direction using Langevin dynamics to output the vehicle joint trajectory.

[0060] It should be noted that the dynamic risk potential field is formed by the superposition of Gaussian repulsive forces generated by surrounding semantic neighbor vehicles, and the collision risk is dynamically quantified based on the vehicle prediction position of the initial joint trajectory.

[0061] In practice, Langevin's dynamic iterative optimization is as follows: starting from the initial predicted trajectory point, the trajectory point is updated in the opposite direction of the dynamic risk potential energy gradient, and random noise is added to complete the trajectory fine-tuning.

[0062] In practice, trajectory optimization is formulated as an energy minimization problem. Using the initial predicted trajectory points... (from) Starting from this point, the Langevin dynamics equations are used to iteratively update the potential energy gradient in the opposite direction: in, To update the step size parameter, For random noise, The risk gradient of the potential field.

[0063] Through the aforementioned iterative mechanism, the risk gradient explicitly "pushes" the target vehicle's future candidate trajectory away from the high-risk collision zone, suppressing and correcting previously unreasonable predictions. The final output is a high-fidelity joint vehicle trajectory that retains both statistical accuracy (reasonable behavioral intent) and strict adherence to physical safety constraints (no collision).

[0064] By constructing the dynamic risk potential field as a superposition of Gaussian repulsive forces generated by surrounding semantic neighbor vehicles, and dynamically quantifying collision risk based on the predicted vehicle positions of the initial joint trajectory, this embodiment overcomes the limitations of static or imprecise construction in traditional risk potential fields. Specifically, each semantic neighbor vehicle is considered a potential collision source, and its influence range and intensity are modeled using a Gaussian function, making the risk potential field spatially smooth and localized. By superimposing these Gaussian repulsive forces, the potential threats of all neighbor vehicles to the target vehicle can be comprehensively considered, forming a comprehensive and refined risk distribution. More importantly, this potential field is dynamically updated based on the predicted vehicle positions, meaning it can reflect the evolution of the traffic scene in real time, thus providing a highly accurate and responsive energy function for subsequent Langevin dynamics optimization. This enables the trajectory optimization process to effectively avoid potential collision areas, generating safer joint trajectories that better reflect actual traffic flow, significantly improving the reliability and safety of the predicted trajectory.

[0065] In the Langevin dynamics iterative optimization process, not only is the energy gradient of the dynamic risk potential field used to guide trajectory points towards the low-risk region in the opposite direction, but more importantly, random noise is introduced to fine-tune the trajectory. This combination enables the optimization process to effectively avoid getting trapped in local optima, thereby exploring a wider range and more diverse range of safe trajectories. The introduction of random noise allows the generated joint vehicle trajectory to maintain low risk while better addressing the inherent uncertainties in real-world traffic scenarios, improving the robustness and adaptability of the predicted trajectory, and providing a more reliable decision-making basis for autonomous driving systems.

[0066] This embodiment includes the following technologies: A query-centralized joint prediction framework decoupled from spatiotemporal representations is proposed: This framework decouples complex traffic scenarios into "vehicle-centric dynamic behavioral features" and "map-centric static structural constraints." It overcomes the spectral bias of multilayer perceptrons by using Fourier features and effectively addresses the heterogeneity problem of multi-agent interactions in complex scenarios by utilizing a bidirectional cross-attention mechanism.

[0067] A causal semantic interaction mechanism based on backdoor adjustment innovatively introduces causal inference theory. A semantic directed graph is constructed in the Frenet coordinate system, and potential confounding factors in multi-agent interactions are simulated through variational inference. The backdoor adjustment mechanism is used to eliminate interference from environmental confounding factors, and the true causal attention weights are calculated, significantly improving the model's robustness in out-of-distribution (OOD) scenarios.

[0068] Combining Langevin Dynamics' physics-guided risk reconstruction mechanism: To address the lack of safety guarantees in purely data-driven models, a dynamic risk potential field based on semantic neighbors is constructed. The predicted vehicle position is treated as a state point in the potential field, and Langevin Dynamics is used as a lightweight iterative optimizer to explicitly apply physical safety constraints to the predicted trajectory, ensuring the physical feasibility and collision-free nature of the final output trajectory from the source.

[0069] This embodiment obtains a comprehensive contextual representation through spatiotemporal feature decoupling encoding and constructs a vehicle semantic topology map in the Frenet coordinate system. A backdoor adjustment mechanism is used for causal semantic interaction modeling, effectively eliminating environmental confounding factors and obtaining true causal attention weights. Based on this, the joint decoder generates a multimodal initial joint trajectory and combines it with a dynamic risk potential field as an energy function. Through Langevin dynamics iterative optimization, the output vehicle joint trajectory achieves improved accuracy, physical feasibility, and safety in complex traffic scenarios, thereby enhancing the performance of the autonomous driving system.

[0070] Reference Figure 2 , Figure 2 This is a structural block diagram of the first embodiment of the vehicle joint trajectory prediction system of this application.

[0071] like Figure 2 As shown, the vehicle joint trajectory prediction system proposed in this application includes: Data acquisition module 10 is used to acquire raw traffic scene data, perform spatiotemporal feature decoupling encoding on the data, and obtain a contextual representation that integrates vehicle dynamic behavior and map static topology; Topology module 20 is used to construct a vehicle semantic topology map in the Frenet coordinate system based on the context representation, and to complete causal semantic interaction modeling through a backdoor adjustment mechanism, thereby eliminating environmental confounding factors and obtaining causal attention weights. The dynamic risk potential field module 30 is used to generate a multimodal initial joint trajectory based on the causal attention weights using a joint decoder, and to construct a dynamic risk potential field based on the predicted positions of semantic neighbor vehicles. Output module 40 is used to take the dynamic risk potential field as an energy function, take the multimodal initial joint trajectory as the starting point, and iteratively optimize along the risk gradient direction through Langevin dynamics to output the vehicle joint trajectory.

[0072] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solution of this application. In specific applications, those skilled in the art can make settings as needed, and this application does not impose any restrictions on this.

[0073] This embodiment eliminates environmental confounding factors through causal semantic interaction modeling and constructs a dynamic risk potential field for iterative optimization. This improves prediction accuracy, physical consistency, and robustness in complex traffic environments. At the same time, it explicitly mines the causal dependencies of vehicle behavior and integrates the physical safety constraints of risk perception.

[0074] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this application. In practical applications, those skilled in the art can select some or all of it to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0075] In addition, for technical details not described in detail in this embodiment, please refer to the vehicle joint trajectory prediction method provided in any embodiment of this application, which will not be repeated here.

[0076] Furthermore, it should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0077] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0078] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of this application. The above are only preferred embodiments of this application and do not limit the patent scope of this application. All equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for predicting joint vehicle trajectories, characterized in that, include: Obtain raw traffic scene data, perform spatiotemporal feature decoupling encoding on the data, and obtain a contextual representation that integrates vehicle dynamic behavior and map static topology; Based on the aforementioned context representation, a vehicle semantic topology map is constructed in the Frenet coordinate system, and causal semantic interaction modeling is completed through a backdoor adjustment mechanism to eliminate interference from environmental confounding factors and obtain causal attention weights. Based on the causal attention weights, a joint decoder is used to generate a multimodal initial joint trajectory, and a dynamic risk potential field is constructed based on the predicted positions of semantic neighbor vehicles. Using the dynamic risk potential field as an energy function, and starting from the initial multimodal joint trajectory, the vehicle joint trajectory is output through iterative optimization along the risk gradient direction using Langevin dynamics.

2. The method according to claim 1, characterized in that, The spatiotemporal feature decoupling encoding specifically involves: embedding Fourier features into the vehicle's continuous kinematic state vector; independently extracting the vehicle's historical trajectory temporal features and vectorized map topological features through a two-stream architecture; and then fusing them through a bidirectional cross-attention mechanism to obtain the context representation.

3. The method according to claim 2, characterized in that, Fourier feature embedding involves mapping the kinematic state vector of vehicle position, velocity, and heading angle to a high-dimensional embedding space through a learnable linear projection matrix after concatenating the vectors using cosine and sine functions.

4. The method according to claim 1, characterized in that, The vehicle semantic topology graph uses a geometric relationship classifier to determine the front, rear, left, and right semantic spatial relationships between the target vehicle and its neighboring vehicles, and constructs a directed semantic interaction graph based on these semantic spatial relationships.

5. The method according to claim 1, characterized in that, The causal semantic interaction modeling is specifically as follows: the potential confounding variables in vehicle interaction are estimated through variational inference, the confounding bias terms in the conventional attention score are removed based on the backdoor adjustment criterion, and the causal attention weights are calculated.

6. The method according to claim 1, characterized in that, The joint decoder is a joint GRU decoder, which generates scene-level multimodal initial joint trajectories synchronously based on the causal attention weights autoregressively.

7. The method according to claim 1, characterized in that, The dynamic risk potential field is formed by the superposition of Gaussian repulsive forces generated by surrounding semantic neighbor vehicles, and the collision risk is dynamically quantified based on the vehicle prediction position of the initial joint trajectory.

8. The method according to claim 1, characterized in that, Langevin dynamics iterative optimization is as follows: starting from the initial predicted trajectory point, update the trajectory point in the opposite direction of the dynamic risk potential energy gradient, and add random noise to complete trajectory fine-tuning.

9. A vehicle joint trajectory prediction system, characterized in that, include: The data acquisition module is used to acquire raw traffic scene data, perform spatiotemporal feature decoupling encoding on the data, and obtain a contextual representation that integrates vehicle dynamic behavior and map static topology; The topology module is used to construct a vehicle semantic topology map in the Frenet coordinate system based on the context representation, and to complete causal semantic interaction modeling through a backdoor adjustment mechanism, thereby eliminating interference from environmental confounding factors to obtain causal attention weights. The dynamic risk potential field module is used to generate a multimodal initial joint trajectory based on the causal attention weights using a joint decoder, and to construct a dynamic risk potential field based on the predicted positions of semantic neighbor vehicles. The output module is used to take the dynamic risk potential field as an energy function, take the multimodal initial joint trajectory as the starting point, and iteratively optimize it along the risk gradient direction through Langevin dynamics to output the vehicle joint trajectory.

Images