Model decision mechanism analysis and optimization method for graph neural network trajectory prediction
By constructing a multi-layer graph neural network model and backpropagating correlation values, the problem of unclear decision-making process in existing trajectory prediction models is solved, enabling accurate positioning of key elements and model optimization, thereby improving the interpretability and safety of autonomous driving systems.
Patent Information
- Application Number
- CN202511337013.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-18
- Publication Date
- 2026-01-09
AI Technical Summary
Existing trajectory prediction methods based on graph neural networks lack a clear explanation of the model's decision-making process, making it difficult to adapt to complex, dynamic, and heterogeneous interaction scenarios. This can lead to misjudgments by autonomous driving systems in complex scenarios, causing safety hazards.
A multi-layer model for trajectory prediction using graph neural networks is constructed. By quantizing and backpropagating correlation values, the correlation distribution of input feature data is decomposed layer by layer. Weight parameters are adjusted to optimize the model structure and parameters, thereby achieving accurate localization and interpretation of key elements.
This improves the interpretability and reliability of the model, enhancing the safety and reliability of autonomous driving systems in complex scenarios.
Smart Images
Figure CN121301840A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of trajectory prediction, in particular to a model decision mechanism analysis and optimization method for graph neural network trajectory prediction. BACKGROUND
[0002] As one of the core technologies of autonomous driving, trajectory prediction has attracted widespread attention and made great progress in recent years. The current mainstream trajectory prediction method based on graph neural network represents the dynamic and static elements and their interaction in the traffic scene through graph data structure, but this kind of method is essentially a "black box" model, which lacks clear explanation of the model decision process. Existing graph neural network explanation methods are mostly oriented to simple classification or regression tasks, which are difficult to adapt to the complex dynamic and heterogeneous interaction scene in trajectory prediction, and cannot effectively reveal how the model generates the predicted trajectory using the input features. In practical applications, if the model decision mechanism cannot be clearly explained, it may lead to misjudgment of the autonomous driving system in complex scenes, causing safety hazards. Therefore, an interpretable method suitable for graph trajectory prediction task is urgently needed, which can quantify the influence of each input feature on the prediction result, accurately locate the key elements affecting the model decision, and then optimize the model structure and parameters, so as to improve the reliability and safety of the autonomous driving system. SUMMARY
[0003] To solve the problems of the prior art, the present application provides a model decision mechanism analysis and optimization method for graph neural network trajectory prediction, which accurately locates the key elements affecting the model decision, thereby optimizing the model structure and parameters.
[0004] To this end, the technical solution adopted by the present application is as follows: The present application provides a model decision mechanism analysis and optimization method for graph neural network trajectory prediction, which comprises: building a multi-layer graph neural network model for trajectory prediction, inputting the input feature data into the multi-layer graph neural network model, and performing trajectory prediction based on the initial weight parameters obtained by model training to obtain the model output result; quantifying and layer-by-layer decomposing the total correlation value corresponding to the predicted trajectory, and based on the correlation value consistency, propagating the obtained correlation distribution from the model output layer to the model input layer in reverse to obtain the correlation distribution of the input feature data of the input layer; According to the correlation distribution of the input layer, quantitatively analyze the influence of each input feature data on the model output result, and adjust the weight parameters of the input feature data in the model combined with the output result deviation to obtain the optimized multi-layer graph neural network model.
[0005] According to the above scheme, the trajectory prediction-oriented multi-layer graph neural network model specifically obtains driving scene data and processes it into a graph data structure, constructs a driving scene representation data graph based on the processed driving scene data, processes information of the driving scene representation data graph, obtains high-dimensional interaction features of driving scene representation interaction relationships of the driving scene representation data graph as input feature data, and generates a trajectory prediction result based on the input feature data representation.
[0006] According to the above scheme, the driving scene representation data graph specifically takes scene elements as nodes and takes interaction relationships between elements as edges connecting the nodes; wherein the scene elements specifically include dynamic scene elements and static scene elements, the dynamic scene elements include motor vehicles, non-motor vehicles and pedestrians, and the static scene elements include road segments, traffic signal lights and obstacles; wherein the interaction relationships between elements include distance constraints and position relationships; the node features of the driving scene representation data graph include node position coordinates, velocities and categories, and the edge features include node relative positions and interaction intensities.
[0007] According to the above scheme, the total correlation value corresponding to the predicted trajectory is decomposed and quantified layer by layer based on the LRP method; in the decomposition process, the graph structure is divided into multiple sub-graphs, and the node correlation values of different sub-graphs are calculated in a parallel manner.
[0008] According to the above scheme, the correlation distribution of the input feature data decomposed layer by layer specifically starts from the output layer of the graph neural network model, and the node correlation values of the output layer, the hidden layer and the input layer are calculated in turn according to the network level order; for each node, according to the connection strength and the activation state of the lower layer node, the correlation value is proportionally distributed; According to the above scheme, the output layer takes the predicted trajectory coordinates output by the multi-layer graph neural network model as nodes for correlation value calculation; the correlation values of the hidden layer and the input layer are specifically calculated according to the connection weights between nodes in the current layer, the activation values of nodes in the current layer and the correlation values of the previous layer, respectively.
[0009] According to the above scheme, the correlation value consistency specifically includes: in the back propagation process, the correlation values of nodes in the same network level are normalized, and the processed correlation values are verified layer by layer, and after each layer of propagation, it is verified whether the total correlation value is consistent.
[0010] According to the above scheme, analyzing the influence of each input feature on the output result of the model specifically includes: constructing an undirected heterogeneous graph with each input feature data as a node, and quantifying the influence degree of each input feature on the output result of the model in a visual representation manner of node shape size or node color depth.
[0011] According to the above scheme, the undirected heterogeneous graph maps the input layer nodes by type into different geometric shapes, wherein the geometric center coordinates of each node correspond to its relative position in the original driving scene, and the influence degree is quantified based on the calculation of the normalized contribution score of the correlation value of the node.
[0012] The application also provides a model decision mechanism analysis and optimization system for graph neural network trajectory prediction, comprising: A trajectory prediction module is configured to construct a multi-layer graph neural network model for trajectory prediction, input the input feature data into the multi-layer graph neural network model, and perform trajectory prediction based on the initial weight parameters obtained through model training to obtain a model output result. A correlation value decomposition and propagation module is configured to quantify and layer-by-layer decompose the total correlation value corresponding to the predicted trajectory, and based on the correlation value consistency, the obtained correlation distribution is reversely propagated from the model output layer to the model input layer to obtain the correlation distribution of the input feature data of the input layer. An analysis and adjustment module is configured to quantitatively analyze the influence of each input feature on the model output result according to the correlation distribution of the input layer, and adjust the weight parameters of the input feature data in the model in combination with the output result deviation to obtain an optimized multi-layer graph neural network model.
[0013] The application has the following beneficial effects: by constructing a multi-layer graph neural network model for trajectory prediction, the application quantifies the total correlation value corresponding to the predicted trajectory, obtains the correlation distribution of the input feature data of each layer in the model and reversely propagates it, analyzes and adjusts the weight parameters based on the correlation distribution propagated to the input layer, accurately locates the key elements affecting the model decision, improves the explainability of the model, provides a clear direction for the optimization and improvement of the model, and helps to improve the reliability and robustness of the model in actual application scenarios.
[0014] Further, the application checks the total correlation value of each layer during the reverse propagation process based on the correlation value consistency, ensures that the total correlation value of each layer node is consistent with the previous layer, makes the total correlation of the input feature consistent with the model output, and ensures the accuracy and reliability of the model decision mechanism analysis. BRIEF DESCRIPTION OF DRAWINGS
[0015] Figure 1 is a method flow diagram of the model decision mechanism analysis and optimization method for graph neural network trajectory prediction of the embodiment of the application; Figure 2 is a schematic diagram of an undirected heterogeneous graph constructed by the embodiment of the application; Figure 3 is a system structure diagram of the model decision mechanism analysis and optimization system for graph neural network trajectory prediction of the embodiment of the application. DETAILED DESCRIPTION
[0016] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not used to limit the present application.
[0017] To solve the problem that the mechanism of the prediction model cannot be effectively analyzed and the key elements affecting the model decision are not positioned in the prior art, an embodiment of the present application provides a model decision mechanism analysis and optimization method for graph neural network trajectory prediction, as shown in Figure 1 The method comprises the following steps: S1, a multi-layer graph neural network model for trajectory prediction is constructed, input feature data is input into the multi-layer graph neural network model, and trajectory prediction is performed based on the initial weight parameters obtained through model training to obtain a model output result.
[0018] S2, the total correlation value corresponding to the predicted trajectory is quantified and decomposed layer by layer, and the correlation distribution obtained is propagated from the model output layer to the model input layer in reverse based on the correlation value consistency to obtain the correlation distribution of the input feature data of the input layer.
[0019] S3, according to the correlation distribution of the input layer, the influence of each input feature on the model output result is quantitatively analyzed, and the weight parameters of the input feature in the model are adjusted combined with the output result deviation to obtain an optimized multi-layer graph neural network model.
[0020] Specifically, the multi-layer graph neural network model for trajectory prediction specifically acquires driving scene data and processes it into a graph data structure, constructs a driving scene representation data graph based on the processed driving scene data, processes information of the driving scene representation data graph, acquires high-dimensional interaction features of driving scene representation interaction relationships of the driving scene representation data graph as input feature data, and generates a trajectory prediction result based on the input feature data. In addition, in the model training stage, the prediction result and the real trajectory are calculated for loss by means of supervised learning, and the graph neural network parameters are updated by using the gradient back propagation algorithm to realize effective convergence of the model and subsequent explainability analysis.
[0021] Specifically, the driving scene representation data graph takes scene elements as nodes and takes interaction relationships between elements as edges; wherein the scene elements specifically include dynamic scene elements and static scene elements, the dynamic scene elements include motor vehicles, non-motor vehicles and pedestrians, and the static scene elements include road segments, traffic signal lights and obstacles; wherein the interaction relationships between elements include distance constraints and position relationships; the node features of the driving scene representation data graph include node position coordinates, speed and category, and the edge features include node relative position and interaction intensity.
[0022] By constructing a driving scene representation data graph, the structured encoding of scene elements is achieved. By introducing common practices such as heterogeneous target encoding and variable subgraph range, and combining attention mechanism to calculate the interaction weight between nodes, high-dimensional semantic features are extracted by multi-layer aggregation to capture long-distance dependencies, thereby modeling the interaction relationship between heterogeneous scene elements. Finally, the extracted high-dimensional interaction features are fused with the local features of the target vehicle, and combined with scene context information such as road topology and traffic rules, a multimodal predicted trajectory containing different turning possibilities is decoded and generated.
[0023] Preferably, the total relevance value corresponding to the predicted trajectory is decomposed and quantified layer by layer based on the LRP method, where LRP (Layer-wise Relevance Propagation) represents the layer-wise propagation of the relevance score; at the same time, during the decomposition process, the graph structure can be divided into multiple subgraphs, and the node relevance values of different subgraphs can be calculated in parallel.
[0024] Specifically, the correlation distribution of the input feature data is decomposed layer by layer, starting from the output layer of the graph neural network model. The correlation values of the nodes in the hidden layer and the input layer are calculated in sequence according to the network hierarchy. For each node, the correlation value is allocated proportionally according to its connection strength and activation state with the nodes in the lower layer.
[0025] In this embodiment, based on the correlation conservation principle of the LRP method, a propagation system is constructed in which the sum of the correlation values of each layer's nodes is equal to that of the previous layer. This system then outputs the predicted trajectory coordinate set from the multi-layer graph neural network model. (in Indicates the future number The coordinates of the trajectory points at each moment are used as the initial correlation values. The design employs a layer-by-layer reverse propagation path from the output layer to the input layer. Starting from the trajectory prediction point in the output layer, and based on its connection weight ratio with the hidden layer nodes, the correlation value is decomposed and transmitted hierarchically until it reaches the scene element node in the input layer, completing the entire propagation chain. The specific propagation process follows these steps: Definition of the first The correlation value of the layer node The output layer is Input layer For the first Layer nodes Its correlation value is backpropagated to the th equation using the following formula. Layer nodes .in, For nodes To the node The connection weights, For nodes In the an activation value of the layer, and the formula ensures that the correlation value is proportionally distributed according to the connection strength and the activation state.
[0026] Specifically, the correlation distribution obtained is inversely propagated from the model output layer to the model input layer based on correlation value consistency, and the correlation value consistency is specifically that, in the process of reverse propagation, the correlation values of nodes in the same network level are normalized, and the processed correlation values are verified layer by layer, and whether the total correlation value is consistent after each layer is propagated is checked.
[0027] After each layer is inversely propagated, whether the total correlation between the current layer and the previous layer is consistent is verified. First, the correlation value is normalized and adjusted: the total correlation sum of the current layer is calculated , and compared with the total correlation sum of the previous layer , to ensure that the final total correlation sum of the input layer is consistent with the model output correlation value .
[0028] Specifically, analyzing the influence of each input feature on the model output result is specifically that, a directed heterogeneous graph is constructed with each input feature data as a node, and the influence degree of each input feature on the model output result is quantified in a visual representation manner of node shape size or node color depth.
[0029] The directed heterogeneous graph maps the input layer nodes into different geometric shapes according to types, wherein the geometric center coordinates of each node correspond to its relative position in the original driving scene, and the influence degree is quantified based on the normalized contribution score of the correlation value of the node, the score is mapped to a preset color space, and the color depth of the node is positively correlated with its contribution degree. In this embodiment, the directed heterogeneous graph constructed is as shown in Figure 2 .
[0030] In addition, the embodiment of the present application also provides a model decision mechanism analysis and optimization system for graph neural network trajectory prediction, as shown in Figure 3 , the system comprises: a trajectory prediction module, configured to construct a multi-layer graph neural network model for trajectory prediction, input input feature data into the multi-layer graph neural network model, and perform trajectory prediction based on the initial weight parameters obtained by model training to obtain a model output result; a correlation value decomposition and propagation module, configured to quantize and decompose the total correlation value corresponding to the predicted trajectory layer by layer, and inversely propagate the correlation distribution obtained from the model output layer to the model input layer based on correlation value consistency to obtain the correlation distribution of the input feature data of the input layer; An analysis and adjustment module is configured to quantitatively analyze the influence of each input feature on the output result of the model according to the correlation distribution of the input layer, and adjust the weight parameter of the input feature data in the model in combination with the output result deviation, to obtain an optimized multi-layer graph neural network model.
[0031] The various modules or mechanisms of the system are mainly used to implement the various steps of the above-mentioned method embodiments, and will not be described one by one here.
[0032] The application provides a model decision mechanism analysis and optimization method for a graph neural network trajectory prediction task, constructs a multi-layer graph neural network model for trajectory prediction, quantifies the total correlation value corresponding to the predicted trajectory, obtains the input feature data correlation distribution of each layer of the model and performs reverse propagation, analyzes and adjusts the weight parameter based on the correlation distribution propagated to the input layer, accurately locates the key elements affecting the model decision, improves the explainability of the model, provides a clear direction for optimization and improvement of the model, and helps to improve the reliability and robustness of the model in actual application scenarios.
[0033] Further, the application verifies the total correlation value of each layer in the reverse propagation process based on the consistency of the correlation value, ensures that the total correlation value of each layer node is consistent with the previous layer, makes the total correlation of the input feature consistent with the model output, and ensures the accuracy and reliability of the model decision mechanism analysis.
[0034] It should be noted that, according to the needs of implementation, each step / component described in the present application can be split into more steps / components, or two or more steps / components or part of the operation of the steps / components can be combined into a new step / component, to achieve the purpose of the present application.
[0035] The size of the serial number of each step in the above embodiments does not mean the execution order, and the execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.
[0036] It should be understood that those skilled in the art can improve or change according to the above description, and all these improvements and changes should belong to the protection scope of the appended claims of the present application.
Claims
1. A model decision mechanism analysis and optimization method for graph neural network trajectory prediction, characterized in that, The method comprises: A multi-layer graph neural network model oriented to trajectory prediction is constructed, input feature data is input into the multi-layer graph neural network model, and trajectory prediction is performed based on initial weight parameters obtained through model training to obtain model output results; The total correlation value corresponding to the predicted trajectory is quantified and decomposed layer by layer, and the correlation distribution obtained is propagated from the model output layer to the model input layer in reverse to obtain the correlation distribution of the input feature data of the input layer; According to the correlation distribution of the input layer, the influence of each input feature data on the model output result is quantitatively analyzed, and the weight parameters of the input feature data in the model are adjusted in combination with the output result deviation to obtain an optimized multi-layer graph neural network model.
2. The model decision mechanism analysis and optimization method for graph neural network oriented trajectory prediction according to claim 1, characterized in that, The multi-layer graph neural network model oriented to trajectory prediction specifically acquires driving scene data and processes it into a graph data structure, constructs a driving scene representation data graph based on the processed driving scene data, processes information of the driving scene representation data graph, acquires high-dimensional interaction features of driving scene representation interaction relationships of the driving scene representation data graph as input feature data, and generates a trajectory prediction result based on the input feature data.
3. The model decision mechanism analysis and optimization method for graph neural network oriented trajectory prediction according to claim 2, characterized in that, The driving scene representation data graph specifically takes scene elements as nodes and takes interaction relationships between elements as edges connecting the nodes; The scene elements specifically include dynamic scene elements and static scene elements, the dynamic scene elements include motor vehicles, non-motor vehicles and pedestrians, and the static scene elements include road segments, traffic signal lights and obstacles; the interaction relationships between elements include distance constraints and position relationships; the node features of the driving scene representation data graph include node position coordinates, speeds and categories, and the edge features include node relative positions and interaction intensities.
4. The model decision mechanism analysis and optimization method for graph neural network oriented trajectory prediction according to claim 2, characterized in that, The total correlation value corresponding to the predicted trajectory is decomposed and quantified layer by layer based on the LRP method; In the decomposition process, the graph structure is divided into multiple subgraphs, and the node correlation values of different subgraphs are calculated in a parallel manner.
5. The model decision mechanism analysis and optimization method for graph neural network oriented trajectory prediction according to claim 1, characterized in that, The correlation distribution of the input feature data decomposed layer by layer specifically starts from the output layer of the graph neural network model, and the node correlation values of the output layer, the hidden layer and the input layer are calculated in order according to the network level; for each node, the correlation value is proportionally distributed according to the connection strength and the activation state of the lower layer nodes.
6. The model decision mechanism analysis and optimization method for graph neural network oriented trajectory prediction according to claim 5, characterized in that, The output layer takes the predicted trajectory coordinates output by the multi-layer graph neural network model as nodes for correlation value calculation; the correlation values of the hidden layer and the input layer are specifically calculated according to the connection weights between nodes in the current layer, the activation values of the nodes in the current layer and the correlation values of the previous layer, respectively.
7. The model decision mechanism analysis and optimization method for graph neural network oriented trajectory prediction according to claim 1, characterized in that, The correlation value consistency specifically is that in the reverse propagation process, the correlation values of nodes in the same network level are normalized, and the processed correlation values are verified layer by layer, and whether the total correlation value is consistent is checked after each layer propagation.
8. The model decision mechanism analysis and optimization method for graph neural network oriented trajectory prediction according to claim 1, characterized in that, Analyzing the influence of each input feature on the model output result specifically is to construct an undirected heterogeneous graph with each input feature data as a node, and to quantitatively analyze the influence degree of each input feature on the model output result in a visual representation manner of node shape size or node color depth.
9. The model decision mechanism analysis and optimization method for graph neural network oriented trajectory prediction according to claim 8, characterized in that, The undirected heterogeneous graph maps input layer nodes by type into different geometric shapes, wherein the geometric center coordinates of each node correspond to its relative position in the original driving scene, and a normalized contribution score is calculated based on the correlation value of the node to quantify the degree of influence.
10. A model decision mechanism analysis and optimization system for graph neural network trajectory prediction, characterized in that, The system comprises: a trajectory prediction module for constructing a multi-layer graph neural network model for trajectory prediction, inputting input feature data into the multi-layer graph neural network model, and performing trajectory prediction based on initial weight parameters obtained through model training to obtain model output results; a correlation value decomposition and propagation module for quantifying and layer-by-layer decomposing total correlation values corresponding to the predicted trajectory, and performing reverse propagation of the obtained correlation distribution from the model output layer to the model input layer based on correlation value consistency to obtain the correlation distribution of the input feature data of the input layer; an analysis and adjustment module for quantitatively analyzing the influence of each input feature on the model output results according to the correlation distribution of the input layer, and adjusting the weight parameters of the input feature data in the model in combination with the output result deviation to obtain an optimized multi-layer graph neural network model.