Vehicle track prediction system based on local behavior query
By using a hierarchical dynamic modal query and trajectory consistency constraint module, and leveraging a local behavior and context-aware agent motion model, multimodal characteristics are captured, solving the problems of trajectory prediction system dependence on high-precision maps and map missingness, and achieving high-precision and robust prediction under map-less conditions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-11
- Publication Date
- 2026-04-07
AI Technical Summary
Existing trajectory prediction systems rely heavily on high-precision maps, resulting in high deployment costs and failure when maps are missing. Furthermore, map-free solutions exhibit poor prediction performance, making it difficult to guarantee the reliability and environmental adaptability of prediction results.
A hierarchical dynamic modality query module, an agent motion model, and a trajectory consistency constraint module are adopted. Through local behavior query and context awareness, a Gaussian mixture model is used to capture multimodal characteristics, and the spatiotemporal consistency constraint module is used to improve the continuity and stability of prediction.
Lightweight deployment and high-precision prediction were achieved without maps, enhancing environmental robustness, solving the system failure risk and deployment limitations caused by map absence, and improving the continuity and stability of prediction.
Smart Images
Figure SMS_180 
Figure QLYQS_121
Abstract
Description
Technical Field
[0001] This invention relates to the fields of intelligent transportation and autonomous driving technology, and in particular to a vehicle trajectory prediction system based on local behavior queries. Background Technology
[0002] As a key module of autonomous driving systems, trajectory prediction provides accurate perception of the future environment for downstream decision-making. In recent years, deep learning-based trajectory prediction systems have become mainstream, with classic models including those using recurrent neural networks to handle temporal dependencies and graph convolutional networks to model spatial interactions. Capturing long-range dependencies. In scene context encoding, early rasterized image techniques rendered scenes as bird's-eye views and extracted road topology and traffic rules using convolutional neural networks. However, due to high computational cost and significant quantization errors, these techniques were gradually replaced by vectorized high-definition maps. The latter uses lightweight vector data to replace raster data, significantly reducing computational costs and quantization errors, becoming the foundational framework in the current trajectory prediction field. However, map-based trajectory prediction systems heavily rely on the real-time updates and coverage of high-precision maps, leading to high deployment costs for data collection and maintenance. Furthermore, the prediction system completely fails when map data is missing, and this rigid dependence on map integrity poses serious safety risks in dynamic scenes. Recent research has attempted to construct map-free trajectory prediction schemes using only sensor data such as LiDAR and cameras, but experiments show that their prediction performance lags significantly behind auxiliary methods relying on high-precision maps. They also face challenges such as scene consistency degradation due to blurred road boundaries and model interpretability difficulties due to the difficulty in tracing complex interaction logic. This technological contradiction manifests as a trade-off between prediction accuracy and generalization ability. Solutions relying on high-precision maps achieve high-precision prediction capabilities through strong scene priors, but at the cost of high deployment costs and failure due to map gaps. Traditional map-free solutions, while enhancing environmental adaptability, struggle to guarantee the reliability of prediction results due to the lack of explicit scene constraints. Therefore, overcoming this trade-off and constructing a trajectory prediction framework that is lightweight, environmentally robust, and does not rely on costly map priors has become a core challenge urgently needing to be addressed in the current autonomous driving field. Summary of the Invention
[0003] Description of the Invention. To overcome the shortcomings of the prior art, this invention proposes a vehicle trajectory prediction system based on local behavior queries to predict vehicle trajectories without a map.
[0004] To achieve the above objectives, the present invention provides the following technical solution: A vehicle trajectory prediction system based on local behavior query, characterized in that the system performs vehicle trajectory prediction in a map-free environment, and the system includes: a hierarchical dynamic modal query module, which predicts vehicle trajectories through modal... The aim is to estimate the modality-aware query of a target agent in a specific scenario by utilizing local behavior. Use context Modules, embedded in scenarios Modality-aware query Optimization was performed to obtain the spatiotemporal features of the intelligent agent. The agent motion model incorporates the spatiotemporal features of the agent obtained from the hierarchical dynamic modality query process. Conditional Modal Query After concatenation, the inputs are fed into two different MLP layers to obtain the current input of the motion model. and ; and a Gaussian mixture model was used to capture the multimodal characteristics of the target vehicle's motion to obtain the prediction results of the agent's motion model. The spatiotemporal consistency constraint module converts the prediction results of the agent's motion model into... Divide into non-overlapping segments each It was then mapped to a time. And by utilizing depthwise convolutions with residual connections and pointwise convolutions, it promotes... Internal and The feature interactions between them ultimately generate Level interaction representation.
[0005] Preferably, the hierarchical dynamic modality query module first initializes the learnable modality query. Each query is responsible for generating One of the trajectories, This represents the number of feature channels in a modal query; then a common feature channel is applied. Layer Enable modal query Able to interact with local behavioral features Interact to obtain modality-aware queries ,in For target intelligent agent The local behavior dependency is derived by using two multilayer perceptron layers with residual connections. To obtain the behavior embedding: In the formula, and This represents a multi-layer perceptron, from which the behavioral embedding is obtained. It is a three-dimensional tensor. This indicates the number of observed motion states. Represents the hidden layer dimension and local behavior. Embed the output behavior Input to Obtaining behavioral sequence features from layers In the formula, Indicates the dimension of the hidden layer.
[0006] Preferably, The layer includes: In the first branch, embedding the behavior Input to linear operations In, through the expansion coefficient Expand the input dimensions: Subsequently, the obtained representations were applied sequentially using convolutional networks, The activation function is processed by the state-space model kernel: In the formula, This represents a one-dimensional convolutional layer. The kernel function represents the state-space model; in the second branch, the behavior is embedded. Input another linear layer middle: Then, the features from these two parallel branches are multiplied and input into the linear layer. : Preserve the time characteristics of the last step. As a local behavioral feature .
[0007] Preferably, the context Includes: time Modeling unit, using common A layered, stacked framework that transforms input variables into high-dimensional embeddings: In the formula, Indicates vehicle Historical trajectories within the observation timeframe, , ,in, and Represents the weight matrix. and This represents the bias vector. The representation layer normalization operation ultimately yields the features. ,in Let the hidden layer dimension be used; then use... To efficiently capture temporal feature dependencies in a task: Will Network output The last time step serves as the temporal dependency representation of the agent. .
[0008] Preferably, the context It also includes: Spatial attention encoding unit: which transforms the agent's temporal dependency representation through three linear transformations. , and , Representing the hidden layer dimensions, query vectors are generated accordingly. Key vector Sum value vector ;right and Perform the dot product operation, and through Operation to generate attention matrix ;Utilizing attention weights Sum value vector To aggregate agent features: The attention function described above can be expressed as: Then, a multi-head attention mechanism is used to extract features: In the formula, Indicates the number of attention heads. This represents the learnable weight matrix. This represents the output features of the multi-head attention layer; The spatiotemporal features are finally generated by sequentially performing two addition and normalization operations and a feedforward neural network. .
[0009] Preferably, the agent motion model includes: agent spatiotemporal features obtained from the hierarchical dynamic modality query process. Conditional Modal Query The data is then concatenated and subsequently input into two different MLP layers to obtain the current input of the motion model. and , ;Will , The position of the target vehicle at the previous time step is input into the motion model. In the formula, and These respectively indicate that the vehicle is in and Velocity in direction, motion model and Each is composed of parameters and The fully connected layer is implemented; the future vehicle speed is obtained through kinematic formulas. Location at any given time: Vehicle No. The future trajectory of a modality can be represented as: In the formula, This indicates the number of prediction time steps for the model.
[0010] Preferably, the agent motion model further includes: using a Gaussian mixture model to capture the multimodal characteristics of traffic participant motion, wherein the Gaussian mixture model is... Indicates the first Mixed weights for each modality For the first The mean vector of Gaussian components, This is the corresponding covariance matrix, used to characterize prediction uncertainty. The agent's motion model receives agent features. and conditional modal queries As input, the mixed weights are output through a multilayer perceptron. And predict covariance using a lightweight neural network head. .
[0011] Preferably, the spatiotemporal consistency constraint module further includes: from a standard Gaussian distribution Medium sampling noise parameters Adjust the noise level to , The scaling factor; this applies to the noise term. Adding this to the prediction results ultimately generates an enhanced trajectory representation: .
[0012] Through the above technical solution, the present invention has the following technical effects: 1. This invention innovatively proposes a map-free multimodal trajectory prediction framework, which captures the general modal features of a scene by mining local behavioral patterns, while modeling the unique attributes of the target agent, thereby mitigating the negative impact of the lack of a high-definition map layer by layer.
[0013] 2. This invention innovatively designs a trajectory consistency constraint module to improve the continuity and stability of the predicted trajectory. This module employs... The hierarchical interaction represents the capture of local temporal correlations between trajectory segments, and enhances the robustness of the model to noise such as sensor errors and environmental interference by introducing spatial noise simulation auxiliary tasks. Detailed Implementation
[0014] Vehicle trajectory prediction aims to infer the future behavior of a moving target by analyzing the agent's historical trajectory and scene semantic information. The system proposed in this invention mainly consists of three core components: a hierarchical dynamic modality query module, an agent motion model, and a trajectory consistency constraint module. This invention first proposes a hierarchical query paradigm that integrates local behavior representations and scene representations to aggregate relationships between multiple scales. Based on modality-aware and context-aware queries, the proposed framework models the future trajectory of the target agent as a Gaussian mixture model, thereby effectively capturing the multimodal characteristics of the target vehicle's motion. To address potential inconsistencies in map-free trajectory prediction tasks, the proposed model also includes a spatiotemporal consistency constraint module.
[0015] (1) Hierarchical dynamic modal query module Local behavioral representations contain rich spatiotemporal contextual information, effectively enhancing the modeling ability of prediction models for complex traffic scenarios. However, fully exploiting the semantics of local behaviors still faces technical bottlenecks. Notably, local behavioral features exhibit unique advantages in representing general scenario modalities, effectively capturing the typical movement patterns of traffic participants. Therefore, this invention proposes a hierarchical query paradigm, constructing a dual... Architecture (i.e., modality) and context This allows for the aggregation of relationships across multiple scales.
[0016] 1) Modality
[0017] Modal The aim is to estimate the potential future motion modalities of a target agent in a specific scene by leveraging local behavior. This module enables the proposed model to perceive general modal features of the scene, providing crucial modal perception guidance for accurate trajectory prediction. The hierarchical dynamic modal query module first initializes the learnable modal query. Each query is responsible for generating One of the trajectories. Here, This represents the number of feature channels in a modal query. Further, a common... Layer ,Right now Enable modal query Able to interact with local behavioral features Interact to obtain modality-aware queries :
[0018] In order to derive the target intelligent agent Local behavior dependency Given its local behavior This invention first uses two multilayer perceptron layers with residual connections to obtain the behavior embedding:
[0019] In the formula, and This represents a multi-layer perceptron. The resulting behavioral embedding... It is a three-dimensional tensor. This indicates the number of observed motion states. Indicates the dimension of the hidden layer.
[0020] To further obtain behavioral sequence features The output behavior is embedded Input to In the layer:
[0021] In the formula, This indicates the dimension of the hidden layer. Specifically, The model consists of three linear layers, one one-dimensional convolutional layer, and a state-space model kernel.
[0022] In the first branch, the invention embeds behavior. Input to linear operations In, through the expansion coefficient Expand the input dimensions:
[0023] Subsequently, the obtained representations were applied sequentially using convolutional networks, The activation function is processed by the state-space model kernel:
[0024] In the formula, This represents a one-dimensional convolutional layer. Represents the kernel function of the state-space model.
[0025] In the second branch, behavior embedding Input to another linear layer middle:
[0026] Furthermore, the present invention performs multiplication operations on the features of these two parallel branches and then inputs them into the linear layer. :
[0027] Furthermore, the present invention retains the time characteristics of the final step. As a local behavioral feature .
[0028] 2) Context
[0029] This invention provides query features for a target-specific context by focusing on the unique behaviors of each agent in a scene. Therefore, this invention utilizes context. Modules, embedded in scenarios Modality-aware query Optimize this. Specifically, conditional modal queries that simultaneously possess modality awareness and context awareness capabilities. The calculation can be performed in the following ways:
[0030] In the formula, Indicates a common Layer Scene embedding Generate by modeling target-specific features.
[0031] Considering the dynamic and complex nature of driving scenarios, this invention decomposes target-specific feature extraction into two independent processes: time... Modeling and spatial attention coding.
[0032] ①Time Modeling For each intelligent agent, this invention employs a shared approach. A layered, stacked framework that transforms input variables into high-dimensional embeddings:
[0033]
[0034] In the formula, Indicates vehicle Historical trajectories within the observation timeframe. , Here, and Represents the weight matrix. and This represents the bias vector. Representation layer normalization operation. To improve computational efficiency, all agents share the same parameters in the embedded network. The final features obtained... (in (The hidden layer dimension) will be input time. The module performs time-series feature modeling.
[0035] because Its selective mechanism allows it to focus on the most relevant features, which is particularly effective when modeling long-term dependencies in dynamic systems. Considering the temporal characteristics of trajectory states, this invention uses Mamba to efficiently capture temporal feature dependencies in the task:
[0036] In the proposed model, the present invention will Network output The last time step serves as the temporal dependency representation of the agent. .
[0037] ② Spatial attention encoding In the driving environment, the interactions of traffic participants significantly impact their trajectories. Accurate trajectory prediction requires a reasonable modeling of these spatial dependencies. In the proposed model, this invention employs... As a spatial interaction feature extractor, because The attention mechanism is particularly well-suited for distinguishing heterogeneous interactions of varying intensities.
[0038] After acquiring the temporal features of all agents, this invention uses three linear transformations. , and ( Generate query vectors respectively (representing the hidden layer dimension). Key vector
[0039] Sum value vector :
[0040] Furthermore, the present invention addresses... and Perform the dot product operation, and through Operation to generate attention matrix The specific calculation method is as follows:
[0041] In the formula, Help maintain Numerical stability of the function. Furthermore, attention weights are utilized. Sum value vector To aggregate agent features:
[0042] The attention function described above can be expressed as: To model spatial interactions in different subspaces, this invention further extends the method by utilizing a multi-head attention mechanism to extract features: In the formula, Indicates the number of attention heads. This represents the learnable weight matrix. This represents the output features of the multi-head attention layer. Furthermore, Through two additions and normalizations in sequence ( The process involves manipulation and a feedforward neural network to ultimately generate spatiotemporal features. .
[0043] To enhance the model's global social perception capability, this invention incorporates spatiotemporal features. By inputting two self-attention layers and further fusing high-order neighborhood agent embedding information, a representation specific to the agent is finally generated. The modal query mechanism designed in this invention integrates local behavioral data with agent-specific features, enabling the proposed model to achieve accurate and reliable trajectory prediction without relying on high-precision maps.
[0044] (2) Intelligent agent motion model The agent motion model uses the spatiotemporal features of the agent obtained from the hierarchical dynamic modality query process. Conditional Modal Query The data is then concatenated and subsequently input into two different MLP layers to obtain the current input of the motion model. and :
[0045] Furthermore, , The position of the target vehicle at the previous time step is input into the motion model:
[0046] In the formula, and These respectively indicate that the vehicle is in and Velocity in direction, motion model and Each is composed of parameters and The fully connected layer is implemented.
[0047] Furthermore, the kinematic formulas can be used to obtain the future... Location at any given time: Therefore, vehicle number The future trajectory of a modality can be represented as: In the formula, This indicates the prediction time step of the model. To capture the multimodal characteristics of traffic participants' motion, this invention models the future trajectory of the target agent as a Gaussian Mixture Model (GMM): In the formula, Indicates the first Mixed weights for each modality For the first The mean vector of Gaussian components, This is the corresponding covariance matrix, used to characterize prediction uncertainty. The agent's motion model receives agent features. and conditional modal queries As input, the mixed weights are output through a multilayer perceptron. It predicts covariance using a lightweight neural network head (linear layer + Softplus). .
[0048] (3) Trajectory consistency constraint In the previous module, the agent motion model used hidden layer features to predict the distribution of the target agent's future trajectory. However, in the absence of prior map information, the prediction model may exhibit low robustness in maintaining temporal and spatial consistency. To address these issues, this invention introduces a trajectory consistency constraint to enhance the robustness of map-free prediction. This constraint enforces continuity and stability by aligning the predicted trajectories, thereby mitigating the negative impacts of temporal independence and spatial inconsistency.
[0049] 1) Time Consistency Constraints To accurately capture time variations in the trajectory, the designed trajectory consistency constraint module employs... Level interaction representation. Specifically, this invention uses the prediction results of the agent's motion model. Divide into non-overlapping segments each It was then mapped to a time. Building upon this, the present invention utilizes depthwise convolution with residual connections and pointwise convolution to promote... Internal and The feature interactions between them ultimately generate Level interaction representation. The above process can be formalized as follows: Therefore, the final predicted trajectory can be represented as: time-varying, and the designed trajectory consistency constraint module adopts... Level interaction representation. Specifically, this invention uses the prediction results of the agent's motion model. Divide into non-overlapping segments each It was then mapped to a time. Building upon this, the present invention utilizes depthwise convolution with residual connections and pointwise convolution to promote... Internal and The feature interactions between them ultimately generate Level interaction representation. The above process can be formalized as follows:
[0050] 2) Spatial Consistency Constraints This invention designs an auxiliary task that enables the proposed model to simulate the prediction results. The invention learns more robust feature representations based on responses to spatial inconsistencies. Specifically, the invention learns from a standard Gaussian distribution. Medium sampling noise parameters To control the degree of spatial distortion, this invention introduces a scaling factor. Adjust the noise to Furthermore, this noise term Adding this to the prediction results ultimately generates an enhanced trajectory representation:
[0051] The map-free vehicle trajectory prediction framework proposed in this invention effectively solves the problems of high real-time update costs, limited coverage, and failure due to map missingness in map-based methods. While completely eliminating the dependence on high-cost map data, it achieves prediction accuracy comparable to map-based trajectory prediction systems through an implicit scene understanding mechanism, thereby addressing the system failure risk and deployment limitations in map-missing scenarios.
[0052] This invention addresses the trajectory deviation problem caused by the lack of explicit scene priors such as road topology and traffic rules in map-free methods. By fully mining local behavior patterns, this invention implicitly captures general scene features such as road topology and common driving behaviors, effectively alleviating the lack of environmental understanding in map-free scenarios.
[0053] The trajectory consistency constraint module designed in this invention effectively overcomes the trajectory jitter and abrupt changes caused by existing methods due to neglecting local temporal correlations or insufficient noise resistance. This module employs... The hierarchical interaction method captures the spatiotemporal dependencies between trajectory segments and introduces spatial noise simulation to enhance the model's robustness to sensor errors and environmental interference, thereby significantly improving the continuity and stability of the predicted trajectory.
[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Those skilled in the art can modify or make equivalent substitutions to the technical solutions of the present invention based on the concept of the present invention, without departing from the essence and scope of the technical solutions of the present invention.
Claims
1. A vehicle trajectory prediction system based on local behavior query, characterized in that, The system predicts vehicle trajectories in a map-free environment, and the system includes: The hierarchical dynamic modal query module, through modality The aim is to estimate the modality-aware query of a target agent in a specific scenario by utilizing local behavior. Use context Modules, embedded in scenarios Modality-aware query Optimization was performed to obtain the spatiotemporal features of the intelligent agent. ; The agent motion model incorporates the spatiotemporal features of the agent obtained from the hierarchical dynamic modality query process. Conditional Modal Query After concatenation, input two different values respectively. The current input to the motion model is obtained in the layer. and ; and a Gaussian mixture model was used to capture the multimodal characteristics of the target vehicle's motion to obtain the prediction results of the agent's motion model. ; The spatiotemporal consistency constraint module converts the prediction results of the agent's motion model into... Divide into non-overlapping segments each It was then mapped to a time. And by utilizing depthwise convolutions with residual connections and pointwise convolutions, it promotes... Internal and The feature interactions between them ultimately generate Level interaction representation.
2. The vehicle trajectory prediction system based on local behavior query according to claim 1, characterized in that, The hierarchical dynamic modality query module first initializes the learnable modality query. Each query is responsible for generating One of the trajectories, This represents the number of feature channels in a modal query; then a common feature channel is applied. Layer Enable modal query Able to interact with local behavioral features Interact to obtain modality-aware queries ,in For target intelligent agent The local behavioral dependencies can be derived in the following ways: Using two multilayer sensing layers with residual connections To obtain the behavior embedding: In the formula, and This represents a multi-layer perceptron, from which the behavioral embedding is obtained. It is a three-dimensional tensor. This indicates the number of observed motion states. Represents the hidden layer dimension and local behavior. ; Embed the output behavior Input to Obtaining behavioral sequence features from layers In the formula, Indicates the dimension of the hidden layer.
3. The vehicle trajectory prediction system based on local behavior query according to claim 2, characterized in that, The layers include: In the first branch, embed the behavior. Input to linear operations In, through the expansion coefficient Expand the input dimensions: Subsequently, the obtained representations were applied sequentially using convolutional networks, The activation function is processed by the state-space model kernel: In the formula, This represents a one-dimensional convolutional layer. Represents the kernel function of the state-space model; In the second branch, the behavior is embedded. Input another linear layer middle: Then, the features from these two parallel branches are multiplied and input into the linear layer. : Preserve the time characteristics of the last step. As a local behavioral feature .
4. The vehicle trajectory prediction system based on local behavior query according to claim 2, characterized in that, The context Includes: time Modeling unit, using common A layered, stacked framework that transforms input variables into high-dimensional embeddings: In the formula, Indicates vehicle Historical trajectories within the observation timeframe, , ,in, and Represents the weight matrix. and This represents the bias vector. The representation layer normalization operation ultimately yields the features. ,in Let the hidden layer dimension be used; then use... To efficiently capture temporal feature dependencies in a task: Will Network output The last time step serves as the temporal dependency representation of the agent. .
5. A vehicle trajectory prediction system based on local behavior query according to claim 4, characterized in that, The context It also includes: Spatial attention encoding unit: which transforms the agent's temporal dependency representation through three linear transformations. , and , Representing the hidden layer dimensions, query vectors are generated accordingly. Key vector Sum value vector ;right and Perform the dot product operation, and through Operation to generate attention matrix ;Utilizing attention weights Sum value vector To aggregate agent features: The attention function described above can be expressed as: Then, a multi-head attention mechanism is used to extract features: In the formula, Indicates the number of attention heads. This represents the learnable weight matrix. This represents the output features of the multi-head attention layer; The spatiotemporal features are finally generated by sequentially performing two addition and normalization operations and a feedforward neural network. .
6. The vehicle trajectory prediction system based on local behavior query according to claim 5, characterized in that, The agent motion model includes: the agent's spatiotemporal features obtained from the hierarchical dynamic modality query process. Conditional Modal Query The data is then concatenated and subsequently input into two different MLP layers to obtain the current input of the motion model. and , ;Will , The position of the target vehicle at the previous time step is input into the motion model. In the formula, and These respectively indicate that the vehicle is in and Velocity in direction, motion model and Each is composed of parameters and The fully connected layer is implemented; the future vehicle speed is obtained through kinematic formulas. Location at any given time: Vehicle No. The future trajectory of a modality can be represented as: In the formula, This indicates the number of prediction time steps for the model.
7. A vehicle trajectory prediction system based on local behavior query according to claim 6, characterized in that, The intelligent agent motion model also includes: using a Gaussian mixture model to capture the multimodal characteristics of traffic participant motion. The Gaussian mixture model is... Indicates the first Mixed weights for each modality For the first The mean vector of Gaussian components, This is the corresponding covariance matrix, used to characterize prediction uncertainty. The agent's motion model receives agent features. and conditional modal queries As input, the mixed weights are output through a multilayer perceptron. And predict covariance using a lightweight neural network head. .
8. The vehicle trajectory prediction system based on local behavior query according to claim 1, characterized in that, The spatiotemporal consistency constraint module also includes: from the standard Gaussian distribution Medium sampling noise parameters Adjust the noise level to , The scaling factor; this applies to the noise term. Adding this to the prediction results ultimately generates an enhanced trajectory representation: .