Multi-source data fusion driven intelligent vehicle traffic accident scene reconstruction method

Abstract

The application belongs to the technical field of intelligent transportation, and provides a multi-source data fusion driven intelligent vehicle traffic accident scene reconstruction method. In view of the problems of insufficient utilization of multi-source heterogeneous data, low spatio-temporal correlation accuracy and lack of accident mechanism analysis of the existing method, a unified space-time datum of vehicle side, road side and cloud side data is constructed, the accurate alignment and missing compensation of heterogeneous data are realized through an adaptive weighted fusion algorithm, and the key process information of the accident is recovered. Then, an accident cause and effect reasoning model based on a space-time graph network is established, and the accident triggering condition, risk transmission path and evolution law are deeply mined. The application realizes high-precision reconstruction and intelligent analysis of the whole process of the accident, significantly improves the integrity, accuracy and explainability of the reconstruction result, and can provide reliable technical support for traffic accident responsibility determination, active safety prevention and control and automatic driving test.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation technology, and in particular relates to a method for reconstructing intelligent vehicle traffic accident scenarios driven by multi-source data fusion. Background Technology

[0002] Accident scene reconstruction is a crucial technical means for traffic accident investigation, liability determination, vehicle safety assessment, and testing and verification of autonomous driving systems. Its core objective is to accurately reconstruct the vehicle's motion state, changes in the traffic environment, and the evolution of the accident before and after it occurs. Currently, accident scene reconstruction mainly relies on techniques such as event data recorder data analysis, surveillance video parsing, vehicle dynamics inversion, and simulation reproduction. However, most existing accident scene reconstruction technologies depend on a single or limited data source to reconstruct the accident process. For example, methods based on EDR or CAN bus data can obtain parameters such as vehicle speed, acceleration, and braking status, but they are difficult to reflect the surrounding traffic environment at the time of the accident. Methods based on video surveillance can describe the accident process, but they are limited by perspective, target occlusion, and environmental conditions, making it difficult to accurately obtain the vehicle's dynamic state. Due to the incomplete acquisition of accident-related information, the results of accident scene reconstruction suffer from insufficient accuracy and missing details.

[0003] With the development of vehicle-road cooperative technology, some studies have begun to utilize vehicle-side and roadside perception data for joint reconstruction. However, due to differences in data acquisition frequency, time reference, and spatial coordinate system among different devices, accurate matching and unified representation of multi-source data are difficult to achieve. On the one hand, existing technologies often employ simple time matching or offline calibration methods for fusion, which easily introduces spatiotemporal biases, affecting the accuracy and reliability of accident scene reconstruction results. Furthermore, in actual accidents, factors such as equipment failure, communication interruptions, perception blind spots, and environmental obstructions often result in missing or discontinuous accident-related data. Existing technologies have limited capacity to process missing data, making it difficult to fully recover key accident process information, thus affecting the accurate reconstruction of the accident evolution process. On the other hand, existing accident scene reconstruction methods mainly focus on reproducing vehicle trajectories and collision processes, paying insufficient attention to the causes of accidents and risk evolution mechanisms. They lack systematic analysis of the correlation between driving behavior, vehicle status, road environment, and traffic factors, making it difficult to reveal the causal chain of the accident. This limits the application value of accident scene reconstruction results in accident mechanism analysis and safety decision-making. Summary of the Invention

[0004] To address at least one of the technical problems mentioned above, this invention provides a method for reconstructing intelligent vehicle traffic accident scenes driven by multi-source data fusion.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of the present invention provides a method for reconstructing intelligent vehicle traffic accident scenes driven by multi-source data fusion, comprising the following steps: S1. Acquire data throughout the entire process before, during, and after an accident, including vehicle-side data, roadside observation data, and cloud-based observation data; S2. After the data from the multi-source data are synchronized in time and space, the global optimal scenario state estimation result is obtained by fusing the data and extracting the accident scenario element database. S3. Establish the vehicle dynamics model and vehicle state equations; for the collision phase, introduce momentum conservation constraints and collision energy constraints; construct and solve the joint recovery objective function to obtain the trajectory of the entire accident process. S4. Combining the global optimal scenario state estimation results and the entire accident trajectory, perform dynamic reconstruction of the spatiotemporal constrained digital twin accident scenario to obtain a digital twin accident scenario that includes vehicle motion state, traffic participant behavior state, road environment state, and risk propagation state. S5. Based on the reconstructed digital twin accident scene data, a two-layer causal reasoning architecture is constructed. The first layer uses a dynamic Bayesian network to model causal relationships and obtain a structural causal chain. The second layer constructs a Granger causal inference model to identify the temporal causal relationships between various factors in the accident evolution process, determine key triggers, and form an accident risk propagation chain. S6. Calculate the credibility evaluation index based on the established multi-dimensional reconstruction evaluation index. When it is determined that the reconstruction result does not meet the credibility requirements, start the feedback optimization mechanism and feed the optimization result back to S2-S6 for iterative optimization until the credibility evaluation index is met, and output the accident reconstruction result.

[0006] Furthermore, the data collected throughout the entire process before, during, and after an accident includes vehicle-side data, roadside data, and cloud-based data. Vehicle-side data includes vehicle CAN bus data, event data recorder data, GNSS positioning data, IMU inertial measurement data, and sensing data from vehicle-mounted cameras, millimeter-wave radar, and lidar. Roadside data includes traffic monitoring videos, data collected by roadside millimeter-wave radar, roadside lidar, roadside units, and traffic signal control equipment. Cloud-based data includes high-precision map data, traffic flow data, meteorological data, V2X communication data, and historical accident data.

[0007] Furthermore, after acquiring vehicle-side data, roadside observation data, and cloud-based observation data, spatiotemporal synchronization and coordinate unification are performed, including: Using UTC time as a unified time standard, timestamps for various types of data are calibrated and synchronized; Using the high-precision map coordinate system as a reference coordinate system, the vehicle coordinate system, sensor coordinate system, and roadside equipment coordinate system are mapped to a unified spatial coordinate framework through a coordinate transformation model; A spatiotemporal synchronization error function is constructed and optimized to achieve a unified spatiotemporal representation of vehicle operation data, environmental perception data, and traffic operation data.

[0008] The process of fusing multi-source data after spatiotemporal synchronization to obtain the globally optimal scene state estimation result includes: Abstract vehicles, road infrastructure, traffic participants, and environmental elements into a graph structure; Construct the state vector of each node in the graph structure, and combine the state vectors of the nodes in the graph structure with the actual relative observations between nodes obtained from vehicle-mounted, roadside, or cloud data to construct the constrained residuals. By combining the constraint residuals, a convex optimization objective function is constructed, and the global optimal estimate set of the state of all scene nodes is obtained by solving the objective function through nonlinear least squares optimization.

[0009] Furthermore, the joint recovery objective function is expressed as: , in, Represents the observation error term. Represents vehicle dynamics constraints. Represents the collision dynamics constraint terms. and The corresponding weighting coefficients, The first one to be restored t Real-time vehicle status. The first one to be restored t+ Vehicle status at moment 1. For the first t Vehicle status observation and estimation at any time For the first t Vehicle control input at all times This represents the discretized vehicle dynamics state transition function. , This represents the total momentum of the system before and after the collision. This represents the collision kinetic energy loss calculated based on the state to be restored. This represents the kinetic energy loss estimated from the observation data. , , This is the weight matrix corresponding to the residuals.

[0010] Furthermore, the dynamic reconstruction of the spatiotemporally constrained digital twin accident scene by combining the globally optimal scene state estimation results and the entire accident trajectory includes: Based on the scene element database and global optimal state estimation, as well as the obtained accident full-process recovery trajectory, an accident scene state sequence is established according to a unified time index. Subsequently, a dynamic spatiotemporal graph is constructed with each traffic entity as a node and the motion relationship, interaction relationship and collision relationship between entities as edges. The spatiotemporal graph is used to describe the dynamic coupling relationship between each subject, and the interaction relationship is input into the scene state transition model to realize the time-by-time evolution of the accident scene. Furthermore, the spatiotemporal risk field is calculated according to the scene node state and interaction relationship, and finally a continuous digital twin accident scene with four stages is formed: before the accident, the accident trigger, the collision and the post-collision.

[0011] Furthermore, the credibility evaluation index is expressed as follows: , , , , , , in, This is due to time consistency error; This refers to spatial reconstruction error; This is the dynamic consistency error; This is due to causal consistency error; To prevent consistency errors in risk propagation, For the first i The observation time of key events For reconstruction time, For actual measured location, To reconstruct the location, For the first i The state vector of the accident vehicle at each sampling time is used for dynamic consistency verification. For the first i The control input vector of the accident vehicle at each sampling time. The first one calculated based on the accident reconstruction trajectory i Rate of change of vehicle state at each sampling time The theoretical rate of change of state calculated from the S4 vehicle dynamics model. Indicates an accident or incident. This represents nodes in a dynamic Bayesian network excluding accidental event nodes. External, a set of conditional nodes that have a direct or indirect causal relationship with the accident event. This represents the conditional probability of an accident occurring. These are the predicted values ​​from the risk field model. The risk field reference value is calculated based on the actual vehicle condition and environmental observations.

[0012] A second aspect of the present invention provides a computer-readable storage medium.

[0013] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the multi-source data fusion-driven intelligent vehicle traffic accident scene reconstruction method described above.

[0014] A third aspect of the present invention provides a computer device.

[0015] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the multi-source data fusion-driven intelligent vehicle traffic accident scene reconstruction method described above.

[0016] A fourth aspect of the present invention provides a program product.

[0017] A program product, which is a computer program product, includes a computer program that, when executed by a processor, implements the steps in the multi-source data fusion-driven intelligent vehicle traffic accident scene reconstruction method described above.

[0018] Compared with the prior art, the beneficial effects of the present invention are: This invention constructs a vehicle-road-cloud-map optimization and fusion model, uniformly mapping vehicle-side data, roadside perception data, and cloud-based traffic operation data onto a graph structure framework. Through node association constraints and a global optimization mechanism, it achieves collaborative fusion of multi-source heterogeneous data. Compared to traditional single-source data or simple weighted fusion methods, this invention can fully utilize multi-dimensional information such as vehicles, roads, traffic participants, and the environment, effectively reducing the impact of data redundancy and observation errors on accident reconstruction results, improving the spatiotemporal correlation accuracy and scene consistency between multi-source data, and providing a high-quality data foundation for subsequent accident reconstruction.

[0019] This invention constructs a vehicle dynamics-collision dynamics joint constraint recovery model. Based on recovering accident process data using vehicle kinematics and dynamic constraints, it further introduces momentum conservation constraints and energy constraints during the collision phase to jointly optimize and solve the vehicle state changes before and after the collision. Compared to traditional trajectory interpolation methods or single dynamics recovery methods, this invention can effectively recover missing information caused by sensor failure, communication interruption, or monitoring blind spots during the accident process, improving the physical consistency and reconstruction accuracy of trajectory recovery during the collision phase, and achieving continuous reconstruction of the motion state throughout the entire accident process.

[0020] This invention constructs a digital twin accident scene reconstruction model based on spatiotemporal constraints. It maps vehicle operating status, traffic participant behavior, road environment status, and traffic control status to a unified digital twin scene space, and builds a risk propagation field model to achieve continuous dynamic reconstruction of the risk accumulation stage before the accident, the accident triggering stage, the collision occurrence stage, and the post-collision evolution stage. Compared to traditional static accident reproduction methods, this invention can not only reconstruct the accident occurrence process but also characterize the spatiotemporal propagation law of accident risk, achieving synchronous reconstruction of the accident scene, accident process, and risk evolution process, thus improving the completeness and interpretability of accident analysis.

[0021] This invention proposes a two-layer causal inference model combining dynamic Bayesian networks and Granger causal inference. The dynamic Bayesian network is used to mine the probabilistic dependencies between accident-related factors, while Granger causal inference is used to identify the temporal causal relationships between factors during accident evolution. By constructing accident causal chains and risk propagation paths, it enables the analysis of the mechanisms by which multi-dimensional factors such as driving behavior, vehicle status, road environment, traffic flow characteristics, and external environmental factors influence the accident formation process. Compared to traditional accident reconstruction methods that only focus on reproducing the accident process, this invention can further identify accident-inducing factors, extract key risk events, and analyze key influencing factors of accidents, thereby improving the ability to analyze accident mechanisms.

[0022] This invention constructs a reconstruction credibility assessment and self-feedback optimization mechanism. It integrates multiple evaluation indicators, including temporal consistency error, spatial reconstruction error, dynamic consistency error, causal consistency error, and risk propagation consistency error, to establish a unified credibility evaluation model. The evaluation results are then fed back to a graph optimization fusion model, a trajectory recovery model, and a causal inference model for iterative optimization. Compared to existing accident reconstruction methods that lack result verification and quality assessment mechanisms, this invention enables adaptive correction and continuous optimization of accident reconstruction results, improving their reliability and stability.

[0023] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0024] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0025] Figure 1 This is a flowchart of the intelligent vehicle traffic accident scene reconstruction method driven by multi-source data fusion provided in the embodiments of the present invention; Figure 2This is a schematic diagram of vehicle-road-cloud data acquisition and spatiotemporal synchronization provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the vehicle-road cloud map optimization and fusion model provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the joint constraint trajectory recovery model provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of a digital twin accident scene reconstruction model provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the credibility assessment and feedback optimization mechanism provided in the embodiments of the present invention. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0027] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0028] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0029] Existing technologies mostly rely on single data sources or simple data fusion methods for accident reconstruction, making it difficult to fully utilize multi-source heterogeneous data generated from vehicle-side, roadside, and cloud platforms. This results in incomplete accident information and limited reconstruction accuracy. Furthermore, inconsistencies in time references, spatial coordinates, and data gaps exist between different data sources, hindering accurate reconstruction of the entire accident process. In addition, existing methods primarily focus on reproducing accident phenomena, lacking the ability to analyze the causes and risk evolution of accidents, making it difficult to reveal causal relationships. Therefore, this invention aims to solve the technical problems of high-precision spatiotemporal fusion of multi-source heterogeneous data from vehicles, roads, and the cloud, recovery of key accident process information, and causal reasoning, achieving high-precision reconstruction of accident scenes and intelligent analysis of accident mechanisms, thereby improving the completeness, accuracy, and interpretability of accident reconstruction results.

[0030] like Figure 1 As shown, this embodiment provides a method for reconstructing intelligent vehicle traffic accident scenes driven by multi-source data fusion, including the following steps: S1: Acquire data throughout the entire process before, during, and after an accident; In this embodiment, the constructed vehicle-road-cloud collaborative data acquisition system includes vehicle-side data such as vehicle CAN bus data, event data recorder (EDR) data, GNSS positioning data, IMU inertial measurement data, and perception data from vehicle-mounted cameras, millimeter-wave radar, and lidar; roadside data includes traffic monitoring videos, data collected by roadside millimeter-wave radar, roadside lidar, roadside units (RSU), and traffic signal control equipment; and cloud-based data includes high-precision map data, traffic flow data, meteorological data, V2X communication data, and historical accident data.

[0031] Establish a unified vehicle state vector: , in: , These are the vehicle's position coordinates in the global coordinate system; , These are the vehicle's longitudinal speed and lateral speed, respectively. , These are the vehicle's longitudinal acceleration and lateral acceleration, respectively. This refers to the vehicle's heading angle; This represents the vehicle's yaw rate.

[0032] The vehicle-side data is defined as follows: , Roadside observation data are defined as: , Cloud-based observation data is defined as: , in: Indicates the time of the accident. t The state vector; This represents the vehicle-side observation data vector, including data acquired from CAN, EDR, GNSS, IMU, and onboard environmental perception equipment; This represents the roadside observation data vector, including data acquired by roadside cameras, millimeter-wave radar, lidar, and RSU; This represents a vector of cloud-based observation data, including high-precision maps, traffic flow, meteorological environment, and V2X data after unified encoding. , and These are the mapping functions for vehicle-side, roadside, and cloud-based observations, respectively. , and Noise was observed from the vehicle side, roadside, and cloud side, respectively.

[0033] S2: Perform spatiotemporal synchronization and coordinate unification on the acquired vehicle-side data, roadside observation data, and cloud-based observation data; Since the data sampling frequencies and time bases of vehicle-mounted equipment, roadside equipment, and cloud platforms are different, this invention establishes a unified spatiotemporal correlation mechanism to achieve accurate alignment of multi-source data.

[0034] like Figure 2 As shown, the specific steps include the following: S201: Using UTC time as a unified time standard, calibrate and synchronize various data timestamps, as shown below: , in, In the formula: This is the original timestamp; The corrected timestamp; , , These are the time drift compensation parameters.

[0035] S202: Using the high-precision map coordinate system as the reference coordinate system, the vehicle coordinate system, sensor coordinate system and roadside equipment coordinate system are mapped to a unified spatial coordinate framework through a coordinate transformation model; The unified coordinate transformation model is represented as: , in: , These are coordinates in the local coordinate system. Global coordinate system coordinates; It is a rotation matrix; It is a translation vector.

[0036] S203: Construct a spatiotemporal synchronization error function and optimize its solution to achieve a unified spatiotemporal representation of vehicle operation data, environmental perception data, and traffic operation data; In this embodiment, a spatiotemporal synchronization error function is constructed by combining time error compensation and spatial registration methods: , in: The total spatiotemporal registration error of multi-source data from vehicle-road-cloud; N The number of associated targets or matched observation pairs used for spatiotemporal matching; For the first i The positions of the associated targets are transformed to the global coordinate system under vehicle-side observation; The position of the same target in the roadside observation transformed to the global coordinate system; For the first i Vehicle-side correction timestamps corresponding to each associated target; For the same target, the roadside correction timestamp; The weighting coefficient of the temporal residual relative to the spatial residual.

[0037] The optimization objective is: , in, To be the optimal rotation matrix, The optimal translation vector; Based on the above methods, a unified spatiotemporal representation of vehicle operation data, environmental perception data, and traffic operation data can be achieved.

[0038] S3: After the data from multiple sources are synchronized in time and space, the global optimal scenario state estimation result is obtained by fusing the data and extracting the accident scenario element database. After completing the spatiotemporal synchronization of multi-source data, to address the issues of observation errors, information redundancy, and inconsistent correlations among vehicle-mounted sensors, roadside sensing devices, and cloud data, this invention constructs a vehicle-road-cloud-map optimization fusion model to achieve consistent global state estimation for accident scenarios, such as... Figure 3 As shown, the specific steps include the following: Abstracting vehicles, road infrastructure, traffic participants, and environmental elements into a graph structure: , in, Represents a set of nodes; Represents a set of edges, with nodes defined as follows: ,in: For the accident vehicle node; For road infrastructure nodes; For traffic participant nodes; For environment nodes.

[0039] The state of each node is represented as follows: , in, For the first i The state vector of each scene entity node; This represents the vertical and horizontal positions of the node in the unified global coordinate system; For node speed; For nodal acceleration; The heading angle of the node.

[0040] Constrained residuals: , in, For nodes i With nodes j Between ij The residuals of the constrained edges; The actual relative observations between nodes are obtained from vehicle-mounted, roadside, or cloud data. An observation model for predicting relative observations based on node status; This represents the state vector of the nodes at both ends of the constraint edge.

[0041] Construct a graph to optimize the objective function: , in, , For the first ij The covariance matrix of the observation error of each constrained edge. This is the corresponding information matrix.

[0042] Solving using nonlinear least squares optimization: ,in, This is the globally optimal set of estimates for the states of all scene nodes.

[0043] Integrating the globally optimal node states obtained from graph optimization The final accident scene element database is constructed by jointly extracting static road / environmental information from S1 and traffic control information after unified spatiotemporal mapping from S2, as shown below: , in, ,in, It indicates the status of the vehicle involved in the accident, including its position, speed, acceleration, heading angle, and yaw rate; It indicates the status of other traffic participants, including their position, speed, category, and relative motion. It represents road elements, including lane markings, road geometry, speed limits, and road surface conditions; It indicates traffic control elements, including traffic light phases, traffic signs, and control instructions; This indicates environmental factors, including weather, visibility, and road surface information.

[0044] S4: Establish the vehicle dynamics model and vehicle state equations; for the collision phase, introduce momentum conservation constraints and collision energy constraints; construct and solve the joint recovery objective function to obtain the trajectory of the entire accident process; To address the issues of data loss, sensor failure, and loss of information at the moment of collision during accidents, this invention constructs a joint constraint recovery model of vehicle dynamics and collision dynamics.

[0045] like Figure 4 As shown, the specific steps include the following: First, establish the vehicle dynamics model, which specifically includes: Vehicle longitudinal dynamics model: , Vehicle lateral dynamics model: , Vehicle yaw dynamics model: , in, For the overall vehicle weight; The moment of inertia is the rotation about the vertical axis; , These are the distances from the front and rear axles to the center of mass, respectively. , These are the lateral forces of the front and rear wheels, respectively. and These are the first derivatives of the longitudinal and lateral velocities with respect to time, respectively. and These are the longitudinal velocity and lateral velocity at the vehicle's center of gravity, respectively. This refers to the yaw rate; This is the yaw acceleration; For the longitudinal resultant force of the vehicle; Then, the vehicle state equation is established: , in, X This is the vehicle's continuous-time state vector; The first derivative of the state vector with respect to time; It is a nonlinear state transition function composed of the longitudinal, lateral, and yaw dynamics of the vehicle; This indicates steering input and driving force input. The steering angle of the front wheels; For longitudinal drive / braking force.

[0046] For the collision phase, a momentum conservation constraint is introduced: , in, and The masses of vehicle 1 and vehicle 2 that collided are respectively; and These are the instantaneous speeds of the two vehicles before the collision; , These are the instantaneous speeds of the two vehicles after the collision; superscript " " indicates the instant before the collision, and the superscript "+" indicates the instant after the collision.

[0047] Further establish collision energy constraints: , Construct a joint recovery objective function: , Among them, the observation error term is The vehicle dynamics constraint terms are The collision dynamics constraint terms are , and The corresponding weighting coefficients, The first one to be restored t Real-time vehicle status. The first one to be restored t+ Vehicle status at moment 1. The first obtained by optimizing the S3 graph t Vehicle status observation and estimation at any time For the first t Vehicle control input at all times This represents the discretized vehicle dynamics state transition function. , This represents the total momentum of the system before and after the collision. This represents the collision kinetic energy loss calculated based on the state to be restored. This represents the kinetic energy loss estimated from collision deformation, EDR, accident site measurements, or other observational data. , , This is the weight matrix corresponding to the residuals.

[0048] Final solution: The entire trajectory of the accident was obtained.

[0049] S5: Based on the global optimal scenario state estimation results and the trajectory data of the entire accident process, construct a digital twin accident scenario dynamic reconstruction model based on spatiotemporal constraints; Scene element database and global optimal state estimation based on S3 The accident scene state sequence is established using the full-process accident recovery trajectory obtained from S4 and a unified time index. Subsequently, a dynamic spatiotemporal graph is constructed with each traffic entity as a node and the motion, interaction, and collision relationships between entities as edges. This spatiotemporal graph describes the dynamic coupling relationships between the entities, and the interaction relationships are input into the scene state transition model to realize the time-by-time evolution of the accident scene. Furthermore, the spatiotemporal risk field is calculated based on the scene node states and interaction relationships, ultimately forming a continuous digital twin accident scene with four stages: before the accident, accident triggering, collision, and after the collision. Figure 5 As shown, the specific steps include the following: First, establish the state space for the accident scenario: , in: A set of vehicle states in the accident; For the set of states of traffic participants; A set of road infrastructure statuses; A set of environmental states; This is a set of traffic control states.

[0050] The set of accident vehicle states is defined as follows: , These represent the vehicle's position coordinates in the global coordinate system, its longitudinal and lateral velocities, its longitudinal and lateral accelerations, its heading angle, and its yaw rate; the traffic participant's state is defined as follows: Then, construct a spatiotemporal graph model of the accident scene: ,in: This represents various entity nodes in the accident scenario. The edge set is defined as: These represent motion relationships, interaction relationships, and collision relationships, respectively. Define the adjacency matrix: ,in, , For nodes i With nodes j distance, This is a vector indicating the interaction relationship.

[0051] Then, the graph structure is input into the scene dynamic evolution model: , in, This is the scene state transition function; Input for driving behavior; This is an input of environmental disturbance.

[0052] To describe the propagation process of accident risks in a scenario, a risk field model is further constructed: , in: The intensity of the spatiotemporal risk field; The distance between the target and the source of risk; Risk weights; This represents the risk diffusion coefficient.

[0053] The entire accident scenario sequence is obtained through continuous time-series updates: , Further divide the accident evolution stages: , These correspond to the following stages: risk accumulation stage; accident triggering stage; collision occurrence stage; and post-collision evolution stage.

[0054] Ultimately, a digital twin accident scenario is formed, encompassing vehicle motion status, traffic participant behavior status, road environment status, and risk propagation status.

[0055] S6: Based on the reconstructed digital twin accident scene data, a two-layer causal reasoning architecture is constructed. The first layer uses a dynamic Bayesian network to model causal relationships and obtain a structural causal chain. The second layer constructs a Granger causal inference model to identify the temporal causal relationships between various factors in the accident evolution process, determine key triggers, and form an accident risk propagation chain. In this embodiment, the first layer uses a dynamic Bayesian network to model causal relationships and obtain a structural causal chain. Establishing a causal network: Node set: They represent: driving behavior Vehicle status Road conditions Traffic flow status Environmental factors Accidents and incidents .

[0056] The joint probability is expressed as: , Probability of accident occurrence: , Obtain the structural causal chain: , Dynamic Bayesian networks are mainly used to characterize the conditional probability dependence between variables and the state transition relationship across time. To further quantify the time-series predictive role of candidate factors, Granger causality test is introduced.

[0057] Therefore, a second-layer Granger causal inference model is constructed: For time series: and , Establish a regression model: , in, Indicates time t The explained accident risk variable; in front i The lag term at each moment; This indicates the lag order of the Granger regression model; For the first i Autoregressive coefficient of order; This represents the random error term in the constrained model.

[0058] Introducing variables X back: , in, Indicates candidate cause variables in front j The lag term at each moment; Indicates the corresponding lagged terms The regression coefficient; representing the variables included. X The random error term of the post-model.

[0059] If the following conditions are met: , And if the regression coefficients are significant overall, then it is considered that... right It has Granger time-series prediction causal relationships.

[0060] Then we consider the variable It is a variable Granger's reason.

[0061] Further define causal contribution: , in: This represents the Bayesian causal probability. Granger causality strength; , These are the weighting coefficients.

[0062] The key trigger was finally identified: , A chain of accident risk propagation is formed: driving behavior → abnormal vehicle condition → risk accumulation → accident triggering → collision.

[0063] S7: Establish a multi-dimensional reconstruction evaluation index system, construct a comprehensive credibility evaluation index, and when the reconstruction result is determined not to meet the credibility requirements, start the feedback optimization mechanism, feed the optimization result back to step S3 for iterative optimization, until the credibility evaluation index is met, and output the accident reconstruction result.

[0064] To improve the reliability of accident reconstruction results and the accuracy of causal analysis, this invention constructs a reconstruction credibility assessment and self-feedback optimization model to conduct quality evaluation and iterative optimization of the entire accident reconstruction process.

[0065] like Figure 6 As shown, a multi-dimensional reconstruction evaluation index system is first established: , in: This is due to time consistency error; This refers to spatial reconstruction error; This is the dynamic consistency error; This is due to causal consistency error; This is due to the consistency error in risk propagation.

[0066] Time consistency error is defined as: , in, For the first i The observation time of key events For reconstruction time.

[0067] Spatial reconstruction error is defined as: , in, For actual measured location, To reconstruct the location.

[0068] The dynamic consistency error is defined as: , in, For the first i The state vector of the accident vehicle at each sampling time is used for dynamic consistency verification. For the first i The control input vector of the accident vehicle at each sampling time. The first one calculated based on the accident reconstruction trajectory i Rate of change of vehicle state at each sampling time The theoretical rate of change of state is calculated from the vehicle dynamics model in step S4.

[0069] The causal consistency error is defined as: , in, Indicates an accident or incident. This represents nodes in a dynamic Bayesian network excluding accidental event nodes. External, a set of conditional nodes that have a direct or indirect causal relationship with the accident event. This represents the conditional probability of the accident occurring.

[0070] The risk propagation consistency error is defined as: , in, These are the predicted values ​​from the S5 risk field model. The risk field reference value is calculated based on the actual vehicle condition and environmental observations.

[0071] Further construct a comprehensive credibility evaluation index: , in: , These are the corresponding weighting coefficients. Furthermore, the smaller the value, the smaller the reconstruction error and the higher the reliability.

[0072] When the following conditions are met: , Indicates the threshold of credibility evaluation index The reconstruction result is then deemed to meet the credibility requirements. when: , Then the feedback optimization mechanism will be activated.

[0073] Credibility evaluation index threshold The data is determined using calibration samples with complete accident observation data. Specifically, accident scenarios are reconstructed and calculated for each calibration sample. CRI Samples whose reconstructed trajectory error, collision time error, and collision position error meet the preset engineering tolerance are considered reliable samples. By iterating through candidate thresholds, the threshold that optimizes the overall performance of the correct acceptance rate of reliable samples and the correct rejection rate of unreliable samples is selected as the reliable sample threshold. .

[0074] Define the feedback optimization function: , in: Optimize the error in the S3 plot; This refers to the S4 trajectory recovery error. This is the error in S6 causal inference.

[0075] Update via gradient: , in: For system parameter set; This is the learning rate.

[0076] The optimization results are fed back to: S3 vehicle-road cloud map optimization fusion model, S4 vehicle dynamics-collision dynamics joint recovery model, and S6 dynamic Bayesian network-Granger two-layer causal inference model, ultimately forming a closed-loop reconstruction architecture of fusion → recovery → reconstruction → inference → evaluation → feedback → re-optimization.

[0077] Example 2 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the multi-source data fusion-driven intelligent vehicle traffic accident scene reconstruction method described above.

[0078] Example 3 This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the multi-source data fusion-driven intelligent vehicle traffic accident scene reconstruction method described above.

[0079] Example 4 This embodiment provides a program product, which is a computer program product, including a computer program. When the computer program is executed by a processor, it implements the steps in the multi-source data fusion-driven intelligent vehicle traffic accident scene reconstruction method described above.

[0080] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0081] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0082] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0083] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0084] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0085] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for reconstructing intelligent vehicle traffic accident scenes driven by multi-source data fusion, characterized in that, Includes the following steps: S1. Acquire data throughout the entire process before, during, and after an accident, including vehicle-side data, roadside observation data, and cloud-based observation data; S2. After the data from the multi-source data are synchronized in time and space, the global optimal scenario state estimation result is obtained by fusing the data and extracting the accident scenario element database. S3. Establish the vehicle dynamics model and vehicle state equations; for the collision phase, introduce momentum conservation constraints and collision energy constraints; construct and solve the joint recovery objective function to obtain the trajectory of the entire accident process. S4. Combining the global optimal scenario state estimation results and the entire accident trajectory, perform dynamic reconstruction of the spatiotemporal constrained digital twin accident scenario to obtain a digital twin accident scenario that includes vehicle motion state, traffic participant behavior state, road environment state, and risk propagation state. S5. Based on the reconstructed digital twin accident scene data, a two-layer causal reasoning architecture is constructed. The first layer uses a dynamic Bayesian network to model causal relationships and obtain a structural causal chain. The second layer constructs a Granger causal inference model to identify the temporal causal relationships between various factors during the accident evolution process, determine key triggers, and form an accident risk propagation chain; S6. Calculate the credibility evaluation index based on the established multi-dimensional reconstruction evaluation index. When it is determined that the reconstruction result does not meet the credibility requirements, start the feedback optimization mechanism and feed the optimization result back to S2-S6 for iterative optimization until the credibility evaluation index is met, and output the accident reconstruction result.

2. The method for reconstructing intelligent vehicle traffic accident scenes driven by multi-source data fusion as described in claim 1, characterized in that, The data collected throughout the entire process before, during, and after an accident includes vehicle-side data, roadside data, and cloud data. Vehicle-side data includes vehicle CAN bus data, event data recorder data, GNSS positioning data, IMU inertial measurement data, and sensing data from vehicle-mounted cameras, millimeter-wave radar, and lidar. Roadside data includes traffic monitoring videos, data collected by roadside millimeter-wave radar, roadside lidar, roadside units, and traffic signal control equipment. Cloud data includes high-precision map data, traffic flow data, meteorological data, V2X communication data, and historical accident data.

3. The method for reconstructing intelligent vehicle traffic accident scenes driven by multi-source data fusion as described in claim 1, characterized in that, After acquiring vehicle-side data, roadside observation data, and cloud-based observation data, spatiotemporal synchronization and coordinate unification are performed, including: Using UTC time as a unified time standard, timestamps for various types of data are calibrated and synchronized; Using the high-precision map coordinate system as a reference coordinate system, the vehicle coordinate system, sensor coordinate system, and roadside equipment coordinate system are mapped to a unified spatial coordinate framework through a coordinate transformation model; A spatiotemporal synchronization error function is constructed and optimized to achieve a unified spatiotemporal representation of vehicle operation data, environmental perception data, and traffic operation data.

4. The method for reconstructing intelligent vehicle traffic accident scenes driven by multi-source data fusion as described in claim 1, characterized in that, The process of fusing multi-source data after spatiotemporal synchronization to obtain the globally optimal scene state estimation result includes: Abstract vehicles, road infrastructure, traffic participants, and environmental elements into a graph structure; Construct the state vector of each node in the graph structure, and combine the state vectors of the nodes in the graph structure with the actual relative observations between nodes obtained from vehicle-mounted, roadside, or cloud data to construct the constrained residuals. By combining the constraint residuals, a convex optimization objective function is constructed, and the global optimal estimate set of the state of all scene nodes is obtained by solving the objective function through nonlinear least squares optimization.

5. The method for reconstructing intelligent vehicle traffic accident scenes driven by multi-source data fusion as described in claim 1, characterized in that, The joint recovery objective function is expressed as: ， in, Represents the observation error term. Represents vehicle dynamics constraints. Represents the collision dynamics constraint terms. and The corresponding weighting coefficients, The first one to be restored t Real-time vehicle status. The first one to be restored t+ Vehicle status at moment 1. For the first t Vehicle status observation and estimation at any time For the first t Vehicle control input at all times This represents the discretized vehicle dynamics state transition function. , This represents the total momentum of the system before and after the collision. This represents the collision kinetic energy loss calculated based on the state to be restored. This represents the kinetic energy loss estimated from the observation data. , , This is the weight matrix corresponding to the residuals.

6. The method for reconstructing intelligent vehicle traffic accident scenes driven by multi-source data fusion as described in claim 1, characterized in that, The dynamic reconstruction of the spatiotemporally constrained digital twin accident scene, combining the globally optimal scene state estimation results and the entire accident trajectory, includes: Based on the scene element database and global optimal state estimation, as well as the obtained accident full-process recovery trajectory, an accident scene state sequence is established according to a unified time index. Subsequently, a dynamic spatiotemporal graph is constructed with each traffic entity as a node and the motion relationship, interaction relationship and collision relationship between entities as edges. The spatiotemporal graph is used to describe the dynamic coupling relationship between each subject, and the interaction relationship is input into the scene state transition model to realize the time-by-time evolution of the accident scene. Furthermore, the spatiotemporal risk field is calculated according to the scene node state and interaction relationship, and finally a continuous digital twin accident scene with four stages is formed: before the accident, the accident trigger, the collision and the post-collision.

7. The method for reconstructing intelligent vehicle traffic accident scenes driven by multi-source data fusion as described in claim 1, characterized in that, Credibility evaluation indicators are expressed as follows: ， ， ， ， ， ， in, This is due to time consistency error; This refers to spatial reconstruction error; This is the dynamic consistency error; This is due to causal consistency error; To prevent consistency errors in risk propagation, For the first i The observation time of key events For reconstruction time, For actual measured location, To reconstruct the location, For the first i The state vector of the accident vehicle at each sampling time is used for dynamic consistency verification. For the first i The control input vector of the accident vehicle at each sampling time. The first one calculated based on the accident reconstruction trajectory i Rate of change of vehicle state at each sampling time The theoretical rate of change of state calculated from the vehicle dynamics model. Indicates an accident or incident. This represents nodes in a dynamic Bayesian network excluding accidental events. External, a set of conditional nodes that have a direct or indirect causal relationship with the accident event. This represents the conditional probability of an accident occurring. These are the predicted values ​​from the risk field model. The risk field reference value is calculated based on the actual vehicle condition and environmental observations.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the intelligent vehicle traffic accident scene reconstruction method driven by multi-source data fusion as described in any one of claims 1-7.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the intelligent vehicle traffic accident scene reconstruction method driven by multi-source data fusion as described in any one of claims 1-7.

10. A program product, said program product being a computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps in the intelligent vehicle traffic accident scene reconstruction method driven by multi-source data fusion as described in any one of claims 1-7.

Images