Intelligent analysis method for vehicle and pedestrian collision accident liability

Abstract

The invention relates to the technical field of traffic accident analysis, and discloses an intelligent analysis method for vehicle and pedestrian collision accident liability, and the method comprises the steps: firstly obtaining multi-source heterogeneous accident data, fusing a cross-platform data source through employing a federal learning framework when the data is insufficient, and reconstructing an accident scene three-dimensional coordinate system; and then, based on a multi-modal data fusion result, constructing a traffic participation entity relation topology model by using a graph neural network, and generating an accident dynamic evolution graph. Then, establishing a collision dynamics digital twin model by utilizing a physical engine, extracting a key collision feature vector, and constructing a responsibility probability distribution model based on a generative adversarial network; and optimizing a responsibility judgment strategy by adopting a double-layer reinforcement learning framework, verifying a physical simulation result through a hierarchical verification mechanism, analyzing a responsibility judgment logic chain, and finally outputting a responsibility analysis report with an interpretable label. The method can accurately and intelligently analyze the accident liability, and has good interpretability.

Description

Technical Field

[0001] The present invention relates to the technical field of traffic accident analysis, and particularly to an intelligent analysis method for the liability of vehicle-pedestrian collision accidents. Background Art

[0002] In today's society, with the continuous growth of the automobile ownership and the acceleration of the urbanization process, the interaction between vehicles and pedestrians on the road is becoming increasingly frequent, and vehicle-pedestrian collision accidents are also increasing. Accurately determining the liability of such accidents is crucial for ensuring traffic safety, safeguarding the legitimate rights and interests of all parties, and improving traffic management efficiency. However, the traditional methods for determining accident liability face many difficulties.

[0003] The traditional methods that rely on manual investigation and subjective judgment have obvious drawbacks. The accident scene is complex and changeable, and the investigators may miss key information due to factors such as personal experience and observation angle. For example, in some accident scenes with blocked vision, it is difficult for humans to accurately restore the movement trajectories of vehicles and pedestrians, resulting in a lack of sufficient basis for liability determination. Moreover, different investigators have differences in the understanding and application of traffic regulations, and it is easy to have inconsistent judgment results for the same accident, reducing the fairness and authority of accident handling.

[0004] The existing accident analysis technologies have limitations in data utilization and model construction. Although various sensors and monitoring devices are widely used in the traffic field, multi-source data are often independent of each other and have different formats, making it difficult to effectively integrate. For example, in-vehicle sensor data focuses on the operating state of the vehicle itself, while road monitoring video data mainly focuses on the overall road conditions. There is no effective correlation mechanism between the two, and it is impossible to form a comprehensive and accurate description of the accident scene. At the same time, traditional analysis models usually only consider single factors or simple logical relationships, and cannot fully mine the complex information in accident data. For example, some models only judge the collision liability based on the vehicle speed and collision location, ignoring important factors such as the pedestrian's behavior intention, traffic signal status, and line of sight obstruction, resulting in inaccurate and incomplete analysis results.

[0005] With the increasingly complex traffic environment, new traffic scenes and accident types are constantly emerging, and the traditional methods and existing technologies are increasingly difficult to meet the actual needs. For example, in the context of the gradual popularization of intelligent connected vehicles, the information interaction between vehicles and external devices has a more significant impact on accidents, but the existing analysis means have not been able to effectively incorporate these new factors. Summary of the Invention

[0006] The purpose of the present invention is to provide an intelligent analysis method for the liability of vehicle-pedestrian collision accidents to solve the problems raised in the above background art.

[0007] To achieve the above object, the present invention provides the following technical solution: An intelligent analysis method for vehicle-pedestrian collision accident liability, the method comprising: Step 1: Obtain multi-source heterogeneous accident data, including in-vehicle sensor data, road monitoring video data, lidar point cloud data, and pedestrian mobile terminal positioning data; when the data integrity is insufficient, use a federated learning framework to fuse cross-platform data sources, complete the missing spatio-temporal trajectories through a dynamic weight allocation algorithm, and reconstruct the three-dimensional coordinate system of the accident scene using a spatio-temporal interpolation algorithm; Step 2: Based on the multi-modal data fusion result, use a graph neural network to construct a traffic participant entity relationship topology model, define vehicles, pedestrians, and road facilities as nodes, use motion trajectories, line-of-sight occlusion, and traffic signal status as edge attributes, and combine an attention mechanism to quantify the interaction weights between nodes to generate an accident dynamic evolution map; Step 3: Use a physics engine to establish a collision dynamics digital twin model, input the map data generated in Step 2, simulate the motion trajectory distribution under different collision initial value conditions through Hamiltonian Monte Carlo sampling, and use a variational autoencoder to reduce the dimension of the high-dimensional parameter space to extract key collision feature vectors; Step 4: Based on a generative adversarial network, construct a liability probability distribution model. The generator takes the collision feature vector as input and outputs a virtual liability determination scheme. The discriminator introduces a traffic law knowledge graph as a prior constraint, and iteratively optimizes the liability allocation boundary conditions through adversarial training to generate an accident liability probability matrix with confidence; Step 5: Use a two-layer reinforcement learning framework to optimize the liability determination strategy. The outer agent uses the maximization of regulatory compliance as the reward function to dynamically adjust the liability weight allocation rule; the inner agent optimizes the conflict resolution path based on the Q-learning algorithm, and realizes strategy generalization by storing historical case data in an experience replay pool; Step 6: Design a hierarchical verification mechanism. The first level uses Bayesian hypothesis testing to verify the consistency between the physical simulation results and the original data. The second level uses a symbolic reasoning engine to analyze whether the liability determination logic chain conforms to traffic law provisions, and outputs a final liability analysis report with an interpretability label.

[0008] Preferably, the federated learning framework in Step 1 uses an asynchronous parameter aggregation mechanism. Each data source node locally trains a lightweight trajectory prediction model. The central server protects data features through differential privacy technology, aggregates model parameters using an elastic average algorithm, and uses a gated recurrent unit network to achieve spatio-temporal alignment of cross-modal data.

[0009] Preferably, the graph neural network in step 2 adopts a hybrid architecture of heterogeneous graph attention layer and spatio-temporal convolutional layer. The node embedding vector contains kinematic features such as velocity, acceleration, and heading angle. The edge attribute matrix introduces relative distance, collision time, and visibility index. Key subgraph structure features are extracted through graph pooling operations.

[0010] Preferably, the variational autoencoder in step 3 adopts a conditional latent variable sampling strategy. In the decoder stage, the collision energy conservation equation is introduced as a physical constraint term. The latent space distribution is regularized by a radial basis function kernel to output interpretable latent variable representations in terms of dimensions.

[0011] Preferably, the discriminator of the generative adversarial network in step 4 integrates knowledge graph embedding technology to transform traffic regulation clauses into high-dimensional semantic vectors. The graph attention mechanism is used to calculate the semantic similarity between the generated solution and the regulation clauses, and a contrastive learning loss function is introduced to enhance the clarity of the discriminative boundary.

[0012] Preferably, the two-layer reinforcement learning framework in step 5 adopts a curriculum learning strategy. The outer agent initializes the policy network parameters through imitation learning. The inner agent uses the twin-delayed deep deterministic policy gradient algorithm to optimize the exploration efficiency of the action space, and designs a priority experience sampling mechanism based on the severity of conflicts.

[0013] Preferably, the symbolic reasoning engine in step 6 adopts a method combining first-order logic rules and fuzzy reasoning. The liability determination result is disassembled into atomic propositions, and the integrity of the logical chain is verified through rule template matching. The evidence theory is used to fuse multi-source uncertain information to generate a traceable explanation path.

[0014] Preferably, the heterogeneous graph attention layer uses a multi-head attention mechanism to process the relationships between different types of nodes respectively. The spatio-temporal convolutional layer uses deformable convolutional kernels to capture non-uniform motion patterns, and a memory enhancement module is introduced to store historical interaction state features.

[0015] Preferably, the conditional latent variable sampling strategy uses normalizing flow technology to construct a reversible transformation function, establish a diffeomorphic mapping between the latent space and the observation space, and correct the probability density distribution through the Jacobian determinant to achieve the joint optimization of physical constraints and data-driven.

[0016] Preferably, the fuzzy reasoning method uses interval-type membership functions to quantify the applicability of regulation clauses, designs a rule activation weight assignment algorithm based on conflict factors, and dynamically adjusts the credibility weights of different evidence sources through an evidence synthesis operator.

[0017] Compared with the prior art, the beneficial effects of the present invention are: At the data processing level, by obtaining multi-source heterogeneous accident data and adopting a federated learning framework to fuse cross-platform data sources, the integrity and accuracy of the data are greatly improved. The asynchronous parameter aggregation mechanism in the federated learning framework enables each data source node to train a lightweight trajectory prediction model locally, which not only protects data privacy but also makes full use of the advantages of local data. The central server uses differential privacy technology to protect data features and adopts an elastic averaging algorithm to aggregate model parameters, ensuring the security and reliability of data during the fusion process. The spatio-temporal alignment of cross-modal data achieved by the gated recurrent unit network effectively solves the problems of inconsistency in time and space among different data sources, laying a solid foundation for subsequent accurate analysis. For example, in an actual accident scenario, after the spatio-temporal alignment of in-vehicle sensor data and road surveillance video data, the movement trajectories of vehicles and pedestrians before the accident can be more accurately restored, avoiding misjudgment of liability caused by data deviation.

[0018] In terms of model construction and analysis, the traffic participant entity relationship topology model and the collision dynamics digital twin model constructed based on the multi-modal data fusion results have powerful information mining capabilities. The graph neural network adopts a hybrid architecture of heterogeneous graph attention layers and spatio-temporal convolutional layers, which can comprehensively capture the complex interaction relationships among vehicles, pedestrians, and road facilities, quantify the interaction weights between nodes to generate an accident dynamic evolution map, and clearly display the accident development process. The collision dynamics digital twin model simulates the movement trajectory distribution under different collision initial value conditions through Hamiltonian Monte Carlo sampling, and combines a variational autoencoder to extract key collision feature vectors, accurately grasping the core elements of the accident. This helps to deeply understand the physical mechanism of the accident occurrence and provides a scientific basis for liability determination. Taking a complex intersection collision accident as an example, through these models, the driving speed of the vehicle, the walking route of the pedestrian, and the impact of road facilities on the accident can be accurately analyzed, thus more accurately determining the liability attribution.

[0019] In the optimization of liability determination strategies, the double-layer reinforcement learning framework plays an important role. The outer agent uses the maximization of regulatory compliance as the reward function to dynamically adjust the liability weight allocation rule, ensuring that the liability determination result meets the requirements of traffic regulations. The inner agent optimizes the conflict resolution path based on the Q-learning algorithm and adopts a curriculum learning strategy and a double-delayed deep deterministic policy gradient algorithm to improve the exploration efficiency of the action space. At the same time, the priority experience sampling mechanism based on conflict severity enables the agent to learn strategies for handling severe conflicts faster. This series of optimization measures effectively improve the accuracy and rationality of liability determination. In practical applications, for some accidents with ambiguous liability definitions, this framework can comprehensively consider various factors and make a more fair and objective determination.

[0020] In terms of the verification mechanism and interpretability, the hierarchical verification mechanism ensures the reliability and interpretability of the analysis results. Bayesian hypothesis testing verifies the consistency between the physical simulation results and the original data to ensure the accuracy of model simulation. The symbolic reasoning engine uses a method that combines first-order logic rules and fuzzy reasoning. It decomposes the liability determination results into atomic propositions, verifies the integrity of the logical chain through rule template matching, and uses the evidence theory to fuse multi-source uncertain information to generate a traceable explanation path. This makes the final liability analysis report not only accurate but also able to clearly show the basis for determination and the logical process, facilitating the understanding and acceptance by the parties related to the accident. For example, when dealing with accidents involving disputes among multiple parties, this interpretability can effectively reduce disputes and improve the efficiency of accident handling. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is the working principle diagram of the method of the present invention; Figure 2 is the working principle diagram of data processing in the federated learning framework; Figure 3 is the flowchart of constructing the liability probability of the generative adversarial network; Figure 4 is the step diagram of verification of the symbolic reasoning engine. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0023] Please refer to Figures 1-4 , the present invention provides an intelligent analysis method for the liability of vehicle-pedestrian collision accidents, and its overall implementation solution is as follows: Step 1: Obtain multi-source heterogeneous accident data, including in-vehicle sensor data, road monitoring video data, lidar point cloud data, and pedestrian mobile terminal positioning data. When the data integrity is insufficient, use the federated learning framework to fuse cross-platform data sources. In this framework, each data source node locally trains a lightweight trajectory prediction model. The central server protects data features through differential privacy technology, aggregates model parameters using the elastic averaging algorithm, and achieves spatio-temporal alignment of cross-modal data with the help of a gated recurrent unit network. Complement missing spatio-temporal trajectories through a dynamic weight allocation algorithm, and reconstruct the three-dimensional coordinate system of the accident scene using a spatio-temporal interpolation algorithm.

[0024] Step 2: Based on the multi-modal data fusion results, construct a traffic participant entity relationship topology model using a graph neural network. Define vehicles, pedestrians, and road facilities as nodes, and use motion trajectories, line-of-sight occlusion, and traffic signal states as edge attributes. The graph neural network adopts a hybrid architecture of heterogeneous graph attention layers and spatio-temporal convolutional layers. The node embedding vectors contain kinematic features such as speed, acceleration, and heading angle. The edge attribute matrix introduces relative distance, time to collision, and visibility metrics. Extract key sub-graph structure features through graph pooling operations, and combine the attention mechanism to quantify the interaction weights between nodes to generate an accident dynamic evolution graph.

[0025] Step 3: Use a physics engine to establish a collision dynamics digital twin model and input the graph data generated in Step 2. Simulate the motion trajectory distribution under different collision initial value conditions through Hamiltonian Monte Carlo sampling, and use a variational autoencoder to reduce the dimension of the high-dimensional parameter space. The variational autoencoder adopts a conditional latent variable sampling strategy, introduces the collision energy conservation equation as a physical constraint term in the decoder stage, regularizes the latent space distribution with a radial basis function kernel, outputs a dimensionally interpretable latent variable representation, and extracts key collision feature vectors.

[0026] Step 4: Construct a liability probability distribution model based on a generative adversarial network. The generator takes the collision feature vector as input and outputs a virtual liability determination scheme. The discriminator integrates knowledge graph embedding technology, converts traffic law clauses into high-dimensional semantic vectors, uses a graph attention mechanism to calculate the semantic similarity between the generated scheme and the law clauses, and introduces a contrastive learning loss function to enhance the clarity of the discriminative boundary. The discriminator introduces a traffic law knowledge graph as a prior constraint, and iteratively optimizes the liability allocation boundary conditions through adversarial training to generate an accident liability probability matrix with confidence.

[0027] Step 5: Optimize the liability determination strategy using a two-layer reinforcement learning framework. The outer agent uses the maximization of law compliance as the reward function to dynamically adjust the liability weight allocation rule; the inner agent optimizes the conflict resolution path based on the Q-learning algorithm. The two-layer reinforcement learning framework adopts a curriculum learning strategy. The outer agent initializes the policy network parameters through imitation learning, and the inner agent uses the twin-delayed deep deterministic policy gradient algorithm to optimize the exploration efficiency of the action space and designs a priority experience sampling mechanism based on the severity of the conflict. Achieve policy generalization by storing historical case data in an experience replay pool.

[0028] Step 6: Design a hierarchical verification mechanism. The first level uses Bayesian hypothesis testing to verify the agreement between the physical simulation results and the original data; the second level uses a symbolic reasoning engine to analyze whether the liability determination logic chain conforms to traffic regulations. The symbolic reasoning engine combines first-order logic rules and fuzzy reasoning, decomposes the liability determination result into atomic propositions, verifies the integrity of the logic chain through rule template matching, and uses the evidence theory to fuse multi-source uncertain information to generate a traceable explanation path. Finally, an interpretable final liability analysis report with labels is output.

[0029] The following further describes the implementation of the present invention in combination with Embodiments 1 to 5.

[0030] Embodiment 1: In Step 1, when the data integrity is insufficient, a federated learning framework is used to fuse cross-platform data sources. This federated learning framework adopts an asynchronous parameter aggregation mechanism, and each data source node (such as in-vehicle sensors of different vehicles, various road monitoring devices, etc.) locally trains a lightweight trajectory prediction model. Taking the in-vehicle sensor data source node as an example, when it locally trains the model, based on the collected data such as vehicle speed and acceleration, a machine learning algorithm is used to construct a trajectory prediction model.

[0031] The central server protects data features through differential privacy technology. Differential privacy technology adds noise to the data query or processing result, making it difficult for attackers to infer specific information about the original data from the output result. Assume the original data is , after adding noise the output is , then it satisfies a certain privacy budget constraint, that is , where and are any two adjacent data sets, represents probability. In this way, while ensuring data availability, the privacy of the data is protected.

[0032] The central server uses the elastic averaging algorithm to aggregate model parameters. The elastic averaging algorithm allows dynamic adjustment of the model parameters of different data source nodes during the aggregation process to adapt to the distribution differences of the data. Let the model parameter of the data source node be , and the aggregated parameter be , then , where is the weight dynamically adjusted according to factors such as node data quality and stability, is the number of data source nodes.

[0033] Achieving spatio-temporal alignment of cross-modal data using a gated recurrent unit network. The gated recurrent unit network (GRU) can effectively handle long-term dependencies in time series data. In this scenario, for different modal data such as in-vehicle sensor data and road surveillance video data, the GRU aligns different modal data in time and space based on the timestamps and feature information of the data, making subsequent data fusion more accurate. For example, the vehicle position information collected by the in-vehicle sensor at a certain moment is matched and aligned with the vehicle position information in the road surveillance video at the same moment, providing a reliable data basis for subsequent analysis.

[0034] Embodiment 2: In this embodiment, by optimizing the architecture and parameter settings of the graph neural network, a traffic participant entity relationship topology model is constructed more precisely, generating a graph that can effectively reflect the dynamic evolution of accidents.

[0035] In step 2, the graph neural network adopts a hybrid architecture of a heterogeneous graph attention layer and a spatio-temporal convolutional layer. The heterogeneous graph attention layer uses a multi-head attention mechanism to process the relationships between different types of nodes respectively. For example, for the relationship between vehicle nodes and pedestrian nodes, the multi-head attention mechanism can focus on information such as their movement trajectories and relative distances from multiple different perspectives. Let the attention weight of the -th head in the multi-head attention mechanism be , then the attention feature calculated between node and node through the -th head is , where is the feature vector of node . The final attention feature is the concatenation or weighted sum of the attention features of all heads, which can capture the complex relationships between different nodes more comprehensively.

[0036] The spatio-temporal convolutional layer uses deformable convolutional kernels to capture non-uniform motion patterns. The sampling positions of traditional convolutional kernels are fixed, while deformable convolutional kernels can adaptively adjust the sampling positions according to the characteristics of the input data. When processing the motion trajectory data of vehicles and pedestrians, deformable convolutional kernels can better adapt to non-uniform motion changes. Assuming the offset of the deformable convolutional kernel is , then the convolutional operation at position can be expressed as , where is the input data, is the convolutional kernel weight, and is the number of sampling points.

[0037] The spatio-temporal convolutional layer introduces a memory enhancement module to store historical interaction state features. The memory enhancement module can record the interaction information between nodes at different time steps, such as the relative position changes and speed changes between vehicles and pedestrians at different moments. This helps to better consider the influence of historical factors on the current state when constructing the accident dynamic evolution graph, making the graph more accurately reflect the development process of the accident.

[0038] The node embedding vector contains kinematic features such as speed, acceleration, and heading angle. The edge attribute matrix introduces relative distance, collision time, and visibility index. The key subgraph structure features are extracted through graph pooling operations. Graph pooling operations can perform dimensionality reduction on graph structure data and retain important subgraph structure information. For example, in a complex accident scene graph containing multiple vehicles and pedestrians, the key subgraph closely related to the collision event can be extracted through graph pooling operations, simplifying the subsequent analysis process.

[0039] Example 3: In step 3, the variational autoencoder adopts a conditional latent variable sampling strategy. This strategy uses normalizing flow techniques to construct a reversible transformation function and establish a diffeomorphic mapping between the latent space and the observation space. Let the observation space be , and the latent space be . The reversible transformation function constructed by normalizing flow techniques is , and its inverse transformation is . Through this mapping relationship, data can be processed more flexibly in the latent space.

[0040] The probability density distribution is corrected by the Jacobian determinant. During the transformation process, the Jacobian determinant is used to correct the probability density, that is, , where and are the probability density functions of the observation space and the latent space respectively. This can ensure that the probability distribution characteristics of the data are reasonably transformed before and after the transformation, enabling the data in the latent space to better reflect the characteristics of the observed data.

[0041] The collision energy conservation equation is introduced as a physical constraint term in the decoder stage. The collision energy conservation equation can be expressed as , where is the total energy before the collision, and is the total energy after the collision, including kinetic energy, potential energy, etc. By introducing this equation, during the decoding process, it is ensured that the generated collision feature vectors conform to physical laws, improving the accuracy and reliability of feature extraction.

[0042] The latent space distribution is regularized by a radial basis function kernel, and a potentially variable representation with interpretable output dimensions is obtained. The radial basis function kernel can make the data distribution in the latent space more reasonable, avoiding over-concentration or dispersion. Let the radial basis function kernel be , where and are data points, and is the bandwidth of the kernel function. Through this regularization process, the potentially variable representation of the output has better interpretability, facilitating subsequent analysis and understanding of collision characteristics.

[0043] Example 4: In step 4, the discriminator of the generative adversarial network integrates knowledge graph embedding technology. The traffic regulation clauses are transformed into high-dimensional semantic vectors. For example, the text information in the regulation clauses is transformed into a vector representation through a word vector model (such as Word2Vec, GloVe, etc.). Let the traffic regulation clause be , and the transformed high-dimensional semantic vector be .

[0044] The graph attention mechanism is used to calculate the semantic similarity between the generated plan and the regulation clauses. The graph attention mechanism can focus on different parts of the vectors of the generated plan and the regulation clauses, thus calculating their similarity more accurately. Let the vector representation of the generated plan be , and the similarity score calculated through the graph attention mechanism be , where is the attention weight, and and are the -th and -th dimensions of the vectors and

[0045] respectively.

[0045] The contrastive learning loss function is introduced to enhance the clarity of the discriminative boundary. The contrastive learning loss function can enable the discriminator to better distinguish between real liability determination plans and generated virtual liability determination plans. Suppose the vector of the real plan is , and the vector of the generated plan is , the contrastive learning loss function can be expressed as , where is the similarity score between the real plan and the generated plan, and is the similarity score between the real plan and other samples. By minimizing this loss function, the discriminator can more accurately determine whether the generated plan complies with traffic regulations, and then optimize the boundary conditions of liability assignment to generate a more accurate accident liability probability matrix with confidence.

[0046] Example 5: This embodiment optimizes the double-layer reinforcement learning framework and the symbolic reasoning engine to improve the optimization effect of the liability determination strategy and the accuracy and interpretability of verification.

[0047] In step 5, the double-layer reinforcement learning framework adopts a curriculum learning strategy. The outer agent initializes the policy network parameters through imitation learning. Imitation learning can learn initial policies from existing successful cases or expert experience. For example, extract the liability determination strategy information from a large number of historical cases of judged vehicle-pedestrian collision accidents to initialize the policy network parameters of the outer agent, so that the outer agent has a certain degree of rationality when starting to learn.

[0048] The inner agent uses the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm to optimize the action space exploration efficiency. The Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm reduces the frequency of policy updates, reduces the risk of overestimation, and improves the stability and convergence speed of the algorithm by introducing two delayed update target networks and a policy smoothing mechanism. Let the policy network be , where is the state, is the policy network parameter. The TD3 algorithm optimizes the policy by updating so that the agent can explore the action space more efficiently and find a better conflict resolution path.

[0049] Design a priority experience sampling mechanism based on conflict severity. Prioritize the experiences according to the severity of the conflicts in the accident, and preferentially sample the experience data of severe conflicts for learning. For example, give higher sampling priority to the experience data of serious collision accidents causing casualties, so that the agent can learn the strategies for handling severe conflicts faster and improve the practicality of the liability determination strategy.

[0050] In step 6, the symbolic reasoning engine adopts a method combining first-order logic rules and fuzzy reasoning. Decompose the liability determination result into atomic propositions, and verify the integrity of the logical chain through rule template matching. For example, decompose the liability determination result of "the vehicle was speeding and did not yield to pedestrians, resulting in a collision accident" into atomic propositions such as "the vehicle was speeding", "did not yield to pedestrians", and "resulted in a collision accident", and then perform matching verification according to the pre-set traffic regulation rule templates.

[0051] Use the evidence theory to fuse multi-source uncertain information to generate a traceable explanation path. The evidence theory can handle the uncertainty of multi-source information and fuse evidence from different sources. Assume there are multiple evidence sources , the credibility of the fused evidence is calculated through the synthesis rules in evidence theory (such as Dempster's synthesis rule), and a traceable interpretation path is generated, making the final responsibility analysis report more interpretable, making it easier for relevant personnel to understand the basis and process of responsibility determination.

[0052] Through the above specific implementation methods, the intelligent analysis method of vehicle-pedestrian collision accident responsibility of the present invention can make full use of multi-source heterogeneous data, with the help of advanced models and algorithms, to achieve accurate and intelligent analysis of accident responsibility, and provide explainable results, providing strong support for traffic accident handling.

[0053] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0054] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent analysis method for liability of vehicle-pedestrian collision accidents, characterized by: The following steps are involved: Step 1: Obtain multi-source heterogeneous accident data, including vehicle sensor data, road monitoring video data, lidar point cloud data, and pedestrian mobile terminal positioning data; when the data integrity is insufficient, use the federated learning framework to fuse cross-platform data sources, use the dynamic weight allocation algorithm to complete the missing spatiotemporal trajectories, and use the spatiotemporal interpolation algorithm to reconstruct the three-dimensional coordinate system of the accident scene; Step 2: Based on the results of multimodal data fusion, a graph neural network is used to construct a topological model of the relationship between traffic-related entities. Vehicles, pedestrians, and road facilities are defined as nodes, and motion trajectories, line of sight occlusion, and traffic signal status are used as edge attributes. The interaction weights between nodes are quantified by combining the attention mechanism to generate a dynamic evolution graph of accidents. Step 3: Use the physics engine to build a collision dynamics digital twin model, input the atlas data generated in step 2, simulate the motion trajectory distribution under different collision initial conditions through Hamiltonian Monte Carlo sampling, use variational autoencoder to reduce the dimension of high-dimensional parameter space, and extract key collision feature vectors; Step 4: Construct a liability probability distribution model based on a generative adversarial network. The generator uses the collision feature vector as input and outputs a virtual liability determination scheme. The discriminator introduces the traffic regulations knowledge graph as a priori constraint, iterates and optimizes the liability allocation boundary conditions through adversarial training, and generates an accident liability probability matrix with confidence. Step 5: A two-layer reinforcement learning framework is used to optimize the responsibility determination strategy. The outer layer agent uses maximizing regulatory compliance as the reward function and dynamically adjusts the responsibility weight allocation rules. The inner layer agent optimizes the conflict resolution path based on the Q-learning algorithm and stores historical case data through the experience replay pool to achieve strategy generalization. Step 6: Design a hierarchical verification mechanism. The first level uses Bayesian hypothesis testing to verify the consistency between the physical simulation results and the original data. The second level uses a symbolic reasoning engine to analyze whether the responsibility determination logic chain complies with traffic regulations and outputs a final responsibility analysis report with explainable labels.

2. The intelligent analysis method for vehicle-pedestrian collision accident liability according to claim 1 is characterized by: The federated learning framework described in step 1 adopts an asynchronous parameter aggregation mechanism. Each data source node locally trains a lightweight trajectory prediction model. The central server protects data features through differential privacy technology, uses an elastic averaging algorithm to aggregate model parameters, and uses a gated recurrent unit network to achieve cross-modal data spatiotemporal alignment.

3. The intelligent analysis method for vehicle-pedestrian collision accident liability according to claim 1 is characterized by: The graph neural network described in step 2 adopts a hybrid architecture of heterogeneous graph attention layer and spatiotemporal convolution layer. The node embedding vector contains kinematic features such as speed, acceleration, and heading angle. The edge attribute matrix introduces relative distance, collision time, and field of view visibility indicators. The key subgraph structural features are extracted through graph pooling operations.

4. The intelligent analysis method for vehicle-pedestrian collision accident liability according to claim 1 is characterized by: The variational autoencoder described in step 3 adopts a conditional latent variable sampling strategy, introduces the collision energy conservation equation as a physical constraint term in the decoder stage, and the latent space distribution is regularized by a radial basis function kernel, and the output dimension is interpretable latent variable representation.

5. The intelligent analysis method for vehicle-pedestrian collision accident liability according to claim 1 is characterized by: The discriminator of the generative adversarial network described in step 4 integrates knowledge graph embedding technology to transform traffic regulations into high-dimensional semantic vectors, uses the graph attention mechanism to calculate the semantic similarity between the generated solution and the regulations, and introduces a contrastive learning loss function to enhance the clarity of the discriminant boundary.

6. The intelligent analysis method for vehicle-pedestrian collision accident liability according to claim 1 is characterized by: The two-layer reinforcement learning framework described in step 5 adopts a curriculum learning strategy. The outer agent initializes the policy network parameters through imitation learning, and the inner agent uses a double-delayed deep deterministic policy gradient algorithm to optimize the efficiency of action space exploration, and designs a priority experience sampling mechanism based on conflict severity.

7. The intelligent analysis method for vehicle-pedestrian collision accident liability according to claim 1 is characterized by: The symbolic reasoning engine described in step 6 uses a method that combines first-order logic rules with fuzzy reasoning to decompose the responsibility determination results into atomic propositions, verifies the integrity of the logic chain through rule template matching, and uses evidence theory to fuse multi-source uncertain information to generate a traceable interpretation path.

8. The intelligent analysis method for vehicle-pedestrian collision accident liability according to claim 3 is characterized by: The heterogeneous graph attention layer adopts a multi-head attention mechanism to process the relationships between different types of nodes respectively. The spatiotemporal convolution layer adopts a deformable convolution kernel to capture non-uniform motion patterns, and introduces a memory enhancement module to store historical interaction state features.

9. The intelligent analysis method for vehicle-pedestrian collision accident liability according to claim 4 is characterized by: The conditional latent variable sampling strategy adopts the standardized flow technology to construct a reversible transformation function, establishes a differential homeomorphism mapping between the latent space and the observation space, and corrects the probability density distribution through the Jacobian determinant to achieve the joint optimization of physical constraints and data-driven.

10. The intelligent analysis method for vehicle-pedestrian collision accident liability according to claim 7 is characterized by: The fuzzy reasoning method adopts interval-type membership functions to quantify the applicability of regulatory provisions, designs a rule-activated weight allocation algorithm based on conflict factors, and dynamically adjusts the credibility weights of different evidence sources through evidence synthesis operators.

Images