Intelligent Reconstruction Method of Road Traffic Accidents Based on Automated Pc-Crash Simulation and MLP Algorithm

Abstract

The present invention belongs to the technical field of road traffic, and particularly relates to an intelligent road traffic accident reconstruction method based on automated Pc-Crash simulation and MLP algorithm; it includes Step 1: Establish a simulation model template in Pc-Crash simulation software according to the road traffic accident scene information; Step 2: Design simulation experiments for the parameters that cannot be confirmed at the road accident scene to obtain a simulation data set; Step 3: Construct, optimize, and train a multi-layer perceptron neural network model; Step 4: With the minimization of the stopping position of the accident participants or the deviation of the tire friction marks in the road traffic accident as the core, construct an objective function, use the MLP neural network model to predict the accident result, and use a global optimization algorithm with multi-start search to obtain the optimal solution of the parameters that cannot be confirmed; compare the simulation result with the actual survey data to verify the accuracy of the optimal solution; realize the intelligent reconstruction of the whole process of road traffic accidents, and provide important technical support for liability determination and compensation handling during traffic accident handling.

Description

Technical Field

[0001] The present invention belongs to the technical field of road traffic, and particularly relates to an intelligent road traffic accident reconstruction method based on automated Pc-Crash simulation and MLP algorithm. Background Art

[0002] Analysis of road traffic accidents is one of the key tasks of public security traffic management departments. Its purpose is to reconstruct the accident scene as detailed as possible and analyze the accident causes, so as to provide a scientific basis for accident liability determination and prevention of similar accidents. Traditional accident analysis methods mainly rely on on-site investigation and manual analysis, which not only take a long time but also have limited accuracy. As a professional traffic accident simulation analysis tool, the reliability of PC-Crash in accident reconstruction has been fully verified, and its application in road traffic accident reconstruction analysis has gradually increased in recent years.

[0003] Artur Ziola used Pc-Crash to reconstruct road traffic accidents and compared the visualized 3D simulation with the actual collision video records, fully demonstrating the reliability of the PC-Crash computer simulation software in vehicle collision simulation. The team of Jiang Gongliang from Chongqing Jiaotong University established a simulation model of the road traffic accident process of vehicle frontal collision and rear-end collision based on the Pc-Crash software. The teams of Zhang Yonggang and Huang Haibo established a simulation model of the road traffic accident process of vehicle-pedestrian collision based on the Pc-Crash software. Qin Chuang, Kong Lingshuang, and Tong Xiaobo respectively established a simulation model of the road traffic accident process of two-wheeler collision based on the Pc-Crash software. The above-mentioned existing technologies all compared the simulation results with the data of driving recorders, video surveillance, or event data recorders (EDRs), and the comparison results proved the reliability of PC-Crash in vehicle collision simulation.

[0004] However, when using PC-Crash software for road traffic accident simulation, only forward simulation can be performed, that is, only by setting various parameters such as vehicle speed, collision angle, collision position, types of vehicles involved in the accident, load, road surface gradient, and friction coefficient, the tire friction marks, the stopping positions of accident participants, and the energy consumed in the collision can be obtained through simulation. This current situation is contrary to the needs of public security traffic management and traffic forensic appraisal. They more need to obtain data such as tire friction marks and the stopping positions of accident participants that can be obtained from on-site investigations of traffic accidents, and inversely obtain data such as the initial speed and angular velocity of accident participants before the accident. Currently, existing intelligent traffic accident reconstruction methods based on Pc-Crash all rely on manually adjusting parameters such as the initial relative positions, speeds, postures, and friction coefficients of accident participants in the model multiple times. By comparing the output results of the adjusted model with the on-site investigation data, the initial speed, angular velocity, etc. of accident participants with the best match are obtained through repeated trial and error. On the one hand, this process consumes a large amount of work of accident investigators and forensic appraisers, seriously affecting the efficiency of accident reconstruction analysis and even the work efficiency of the public security traffic management department. On the other hand, this process highly depends on the historical experience of software operators, posing extremely high requirements for the talent reserve of relevant departments. Summary of the Invention

[0005] The purpose of the present invention is to provide an intelligent road traffic accident reconstruction method based on automated Pc-Crash simulation and MLP algorithm. This method realizes the batch execution of collision simulations by constructing an automated control module of Pc-Crash based on the operating system event injection mechanism; at the same time, it innovatively establishes a multi-layer perceptron (MLP) neural network proxy model, and combines intelligent optimization algorithms to achieve rapid convergence of the accident parameter space, thereby realizing the intelligent reconstruction of the entire process of road traffic accidents, providing an important technical means for the public security department to conduct more scientific and fair liability determination and compensation handling in traffic accident handling.

[0006] To achieve the above purpose, the present invention adopts the following technical solutions:

[0007] An intelligent road traffic accident reconstruction method based on automated Pc-Crash simulation and MLP algorithm, comprising the following steps:

[0008] Step 1: Establish a simulation model template in the Pc-Crash simulation software according to the on-site information of the road traffic accident;

[0009] Step 2: Reasonably design simulation experiments within the parameter space for the parameters that cannot be confirmed at the road accident site, and obtain a simulation data set through automated simulation by the automated simulation module according to the designed simulation experiments;

[0010] Step 3: Construct a multi-layer perceptron neural network model, optimize the multi-layer perceptron neural network model, and use the simulation dataset in Step 2 to train the optimized multi-layer perceptron neural network model to obtain the final multi-layer perceptron neural network model;

[0011] Step 4: With the minimization of the deviation of the stopping position or tire friction marks of the accident participants in a road traffic accident as the core, construct an objective function, use the final multi-layer perceptron neural network model in Step 3 to predict the accident result, and obtain the optimal solution of the parameters that cannot be confirmed by a global optimization algorithm with multi-start search; Use the optimal solution of the parameters that cannot be confirmed to conduct a simulation in the PC-Crash simulation software, compare the simulation result with the actual survey data, and verify the accuracy of the optimal solution.

[0012] Further, in Step 1, establishing a simulation model template in the Pc-Crash simulation software according to the road traffic accident scene information means:

[0013] Determine the road conditions at the accident location according to the general overview map of the accident scene taken at the road traffic accident scene and the subsequent on-site surveying or electronic map, and establish a simulation background map or road model in the Pc-Crash simulation software;

[0014] Obtain the information of the accident-involved vehicles according to the road traffic accident scene investigation, and establish an accurate and reliable vehicle model in the Pc-Crash simulation software;

[0015] Analyze the braking marks according to the principles of traceology to obtain the mutation points of the braking mark trajectory, and then obtain the geographical location where the collision occurred; Analyze the vehicle collision relative angle by analyzing the shape of the braking marks and the vehicle deformation characteristics; Establish a suitable relative position of the vehicles in the Pc-Crash simulation software according to the geographical location where the collision occurred and the vehicle collision relative angle;

[0016] Based on video surveillance, driving recorders, and EDR devices, analyze the vehicle speeds and avoidance angular velocities of the accident-involved vehicles, and establish a vehicle motion model in the Pc-Crash simulation software according to the vehicle speeds and avoidance angular velocities.

[0017] Further, the process of the simulation experiment design in Step 2 is as follows:

[0018] For the parameters that cannot be determined at the road accident scene, use the values given at the accident scene as reference values, and set data intervals according to the reference values;

[0019] For the high-dimensional parameter space, use Latin hypercube sampling, and for the low-dimensional parameter space, use a full factorial experimental design to design test points; Among them, high-dimensional parameters refer to data with a dimension greater than 5, and low-dimensional data refer to data with a dimension less than or equal to 5;

[0020] According to the designed test points, the accident process under various parameters is simulated through an automated simulation module, and the tire friction marks and the stopping positions of accident participants are extracted, so as to form a simulation data set of the attributes and behaviors of accident participants and the characteristics of the vehicle and marks after the accident.

[0021] Further, the automated simulation module refers to being embedded in the Pc-Crash simulation software to achieve automated simulation and output simulation results;

[0022] The implementation of the automated simulation module:

[0023] (1) Based on the MATLAB-Java cross-platform interface, the Java AWT Robot class API is called to simulate mouse and keyboard operations; simulating mouse input to achieve point selection operations on the Pc-Crash simulation software, and simulating keyboard input to achieve the input of specific parameters in the Pc-Crash simulation software; furthermore, achieving the save-as of the accident process simulation model template and the accident speed parameter adjustment process;

[0024] (2) Based on the MATLAB-Java cross-platform interface, the Java AWT Robot class API and the MATLAB screen capture and optical character recognition (OCR) functions are called to simulate point selection operations to achieve the simulation run of the adjusted model, and the process status is monitored by recognizing the progress bar value of the image;

[0025] (3) When it is recognized that the simulation process reaches its maximum simulation time, the simulation terminates, and the simulation result file is output by simulating mouse and keyboard operations;

[0026] (4) The output result is obtained by analyzing and reading the simulation result file through Matlab.

[0027] Further, the multi-layer perceptron neural network model includes an input layer, a hidden layer, and an output layer; the ReLU activation function is used in the hidden layer, and the linear activation function is used in the output layer, and the depth and width of the network are dynamically adjusted according to the parameter space dimension.

[0028] Further, optimizing the multi-layer perceptron neural network model means: using the Bayesian optimization algorithm to optimize the hyperparameters of the multi-layer perceptron neural network model; constructing the mapping relationship between the hyperparameters and the model performance through the Gaussian process model, aiming to minimize the mean square error of the validation set, and iteratively searching for the optimal hyperparameter combination; evaluating the generalization ability of the multi-layer perceptron neural network model through K-fold cross-validation.

[0029] Further, in step 4, it is verified that the optimal solution is not the optimal solution by comparing the simulation results with the actual survey data, and the multi-layer perceptron neural network model is further trained by adding simulation test points in step 2 to improve the reliability of the multi-layer perceptron neural network model.

[0030] The beneficial effects of the present invention are as follows:

[0031] (1) The automated simulation module can replace accident investigators and forensic appraisers to perform software operations, forming an automated simulation system for the automobile collision process based on the Pc-Crash simulation software, thereby reducing the workload of accident investigators and forensic appraisers and improving the work efficiency of accident reconstruction analysis and even the public security traffic management department.

[0032] (2) The multi-layer perceptron (MLP) neural network model realizes the intelligent optimization of model parameters and the intelligent reconstruction of the whole process of road traffic accidents, providing an important technical means for the public security department to make more scientific and fair liability determination and compensation handling in traffic accident handling. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flow block diagram of the present invention.

[0034] Figure 2 It is a diagram of the accident process in an embodiment of the present invention.

[0035] Figure 3 It is a surface diagram of the surrogate model in an embodiment of the present invention.

[0036] Figure 4 It is a regression diagnosis diagram of the surrogate model in an embodiment of the present invention.

[0037] Figure 5 It is a comparison diagram of the optimized model simulation and the vehicle stop position at the accident scene of the present invention.

[0038] Figure 6 It is a comparison diagram of the given parameter simulation and the vehicle stop position at the accident scene of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] A method for intelligent reconstruction of road traffic accidents based on automated Pc-Crash simulation and MLP algorithm provided in this embodiment depends on the Pc-Crash simulation software to implement. Before describing the method of this embodiment, an automated simulation module is embedded in the Pc-Crash simulation software, and automated simulation is realized through the automated simulation module to output simulation results, reducing the labor intensity of staff and improving the work efficiency of accident analysis.

[0040] Implementation of the automated simulation module:

[0041] (1) Based on the MATLAB-Java cross-platform interface, the Java AWT Robot class API is called to simulate mouse and keyboard operations. The mouse input is simulated through the mouseMove, mousePress, and mouseRelease functions to perform click operations on the Pc-Crash simulation software, and the keyboard input is simulated through the keyPress and keyRelease functions to input specific parameters in the Pc-Crash simulation software. Furthermore, the save-as operation of the accident process simulation model template and the accident speed parameter adjustment process are realized.

[0042] (2) Based on the MATLAB-Java cross-platform interface, the Java AWT Robot class API and MATLAB screen capture and optical character recognition (OCR) functions are called to simulate click operations to realize the simulation operation of the adjusted model, and the process status is monitored by recognizing the progress bar value of the image.

[0043] (3) When it is recognized that the simulation process reaches its maximum simulation time, the simulation terminates, and the output of the simulation result file is realized by simulating mouse and keyboard operations.

[0044] (4) The output result is obtained by analyzing and reading the simulation result file through Matlab; the output result includes but is not limited to the stop position of the accident participants, the speed, acceleration, and displacement history of the accident participants.

[0045] The automated simulation module of this embodiment can replace accident investigators and forensic appraisers to perform software operations, forming an automated simulation system for vehicle collision processes based on the Pc-Crash simulation software, thereby reducing the workload of accident investigators and forensic appraisers and improving the work efficiency of accident reconstruction analysis and even the public security traffic management department.

[0046] Taking the road traffic accident of "a car hitting a trailer head-on at an intersection" as an example, as Figure 2 shown, the intelligent road traffic accident reconstruction method based on automated Pc-Crash simulation and MLP algorithm described in this embodiment is introduced in detail. As Figure 1 shown, the method includes the following steps:

[0047] Step 1: Establish a simulation model template in the Pc-Crash simulation software according to the road traffic accident scene information.

[0048] Determine the road conditions at the accident site based on the overview map of the accident scene taken at the road traffic accident site and the subsequent on-site surveying and mapping or electronic map, and establish a simulation background map or road model in the Pc-Crash simulation software. In this embodiment, the road conditions at the accident site are determined by the overview map of the accident scene of "a car colliding head-on with a semi-trailer at an intersection" taken by a drone; the taken overview map is used as the background map for the accident process simulation, which is convenient for subsequent comparison of the simulation results with the accident scene.

[0049] Obtain the information of the accident-involved vehicles according to the on-site investigation of the road traffic accident, and establish an accurate and reliable vehicle model in the Pc-Crash simulation software; the vehicle information includes but is not limited to dimensions, wheelbase, tire parameters, and total weight. In this embodiment, the car model is the 2015 Chevrolet Spark M300, the tractor is the 1994 Peterbilt 397, and the semi-trailer is the 2017 Travis S / 96 rear-dump body semi-trailer. The total weight of the semi-trailer and its load is 37 tons; obtain the vehicle information from the website according to the vehicle model, and then establish the vehicle model.

[0050] Analyze the braking marks (morphology, direction, length, etc.) according to the principles of traceology to obtain the mutation points of the braking mark trajectory, and then obtain the geographical location where the collision occurred; analyze the morphology of the braking marks and the deformation characteristics of the vehicle to obtain the relative angle of vehicle collision; establish a suitable relative position of the vehicles in the Pc-Crash simulation software according to the geographical location where the collision occurred and the relative angle of vehicle collision.

[0051] If the braking marks are not obvious or the vehicle deformation characteristics are complex, determine the geographical location where the collision occurred and the relative angle of vehicle collision in combination with the statements of the accident-involved personnel, eyewitnesses, or other debris characteristics.

[0052] In the accident scene investigation of this embodiment, the mutation points of the braking marks are analyzed according to the principles of traceology to determine the geographical location where the collision occurred; the relative angle of vehicle collision is determined based on the morphology of the braking marks and the deformation characteristics of the vehicle; the relative positions of the vehicles in the Pc-Crash simulation software are set according to the obtained parameters, such as Figure 2 The figure shows the simulation model with the relative positions set in the Pc-Crash simulation software, and the background picture in the simulation model is the overview map of the accident scene collected by the drone in the embodiment.

[0053] Based on detection devices such as video surveillance, driving recorders, and EDR, analyze the vehicle speeds and avoidance angular velocities of the accident-involved vehicles, and establish a vehicle motion model in the Pc-Crash simulation software according to the vehicle speeds and avoidance angular velocities.

[0054] If the above-mentioned device does not exist or is damaged, based on the transcripts of the accident participants, the transcripts of the witnesses, the length of the tire friction marks, etc., a rough vehicle speed and avoidance angular velocity are given.

[0055] In the accident scene of this embodiment, there is no video surveillance device, the tow truck is not equipped with a driving recorder and an EDR device, the car is equipped with an EDR device, and the vehicle speed of the car at the time of the accident is recorded. The vehicle speed of the car is the data recorded by the EDR, that is, the initial speed of the car is 33.8 km / h and the angular velocity is 0 rad / s. Considering that both the driver and the passengers of the car suffered fatal injuries and died, the initial speed of the tow truck in the accident process simulation model can only be initially set to 83.6 km / h according to the transcript of the tow truck driver, and the angular velocity is set to 0.2 rad / s.

[0056] Step 2: Reasonably design simulation experiments within the parameter space for the parameters that cannot be confirmed at the road accident scene, and obtain a simulation data set through automated simulation using the automated simulation module according to the designed simulation experiments;

[0057] Due to various reasons such as incomplete or damaged detection equipment, it is difficult to determine multiple parameters such as the vehicle speed and avoidance angular velocity of accident participants before a large number of traffic accidents occur. For the above-mentioned difficult-to-determine parameters, it is necessary to reasonably design simulation experiments within the parameter space so as to find the global optimal matching parameter values subsequently.

[0058] The process of designing the simulation experiment is as follows:

[0059] For the parameters that cannot be determined at the road accident scene, use the values given at the accident scene in Step 1 as reference values, and set data intervals according to the reference values;

[0060] For high-dimensional parameter spaces (dimensions greater than 5), use Latin hypercube sampling, and for low-dimensional parameter spaces (≤5), use full factorial experimental design to design test points;

[0061] According to the designed test points, simulate the accident process for each parameter through the automated simulation module, and extract result data such as tire friction marks and the stopping positions of accident participants, thereby forming a simulation data set of accident participant attributes, behaviors, and vehicle and trace characteristics after the accident.

[0062] In this embodiment, the speed and angular velocity of the trailer are parameters that cannot be determined; taking 83.6 km / h and 0.2 rad / s in Step 1 as reference values, the initial speed range of the trailer is set to 60 to 120 km / h, and the angular velocity range is set to -0.9 to 0.9 rad / s. The full factorial experimental design is used to set the test points, with each parameter taking 13 levels, so a total of 169 groups of simulation test points are determined; the accident process under each parameter is simulated through the automated simulation module, and the result data of the stop positions of the accident participants are extracted, thus forming a simulation data set of the trailer speed, angular velocity, and the vehicle stop position after the accident.

[0063] Step 3: Construct a multi-layer perceptron (MLP) neural network model, optimize the MLP neural network model, and use the simulation data set in Step 2 to train the optimized MLP neural network model to obtain the final MLP neural network model.

[0064] The multi-layer perceptron (MLP) neural network model includes an input layer, a hidden layer, and an output layer; the ReLU activation function is used in the hidden layer to avoid the problem of gradient disappearance, and the linear activation function is used in the output layer to adapt to continuous response variables. The depth and width of the network are dynamically adjusted according to the parameter space dimension, initially set to 1 hidden layer, and the number of neurons in each layer is 4 times the number of input layer variables. The input variables of the input layer are parameters that cannot be confirmed, such as the initial linear speed of collision, angular velocity, and vehicle mass distribution; the output response of the output layer is result data such as tire friction marks, the stop positions of accident participants, and vehicle termination states. The input layer variables in this embodiment are from the simulation data set generated in Step 2, and the output layer response is the simulation result of the PC-Crash simulation software.

[0065] Optimizing the MLP neural network model specifically means: using the Bayesian optimization algorithm to optimize the hyperparameters of the MLP network (such as the number of hidden layers, the number of neurons, etc.); constructing the mapping relationship between the hyperparameters and the model performance through the Gaussian process model, aiming to minimize the mean squared error (MSE) of the validation set, and iteratively searching for the optimal hyperparameter combination. The generalization ability of the surrogate model is evaluated through K-fold cross-validation; the simulation data set is randomly divided into K subsets, and one of the subsets is used as the validation set in turn, and the remaining subsets are used as the training set, and the average performance metrics (such as R 2 , MSE) of the model in K validations are calculated.

[0066] The MLP neural network model is trained using the simulation data set to form a high-precision surrogate model that can replace physical simulation; this model can predict accident results with a response time of milliseconds, significantly reducing the computational cost and providing efficient support for subsequent parameter optimization.

[0067] The optimized MLP neural network model in this embodiment includes 1 hidden layer with 12 neurons. The MLP neural network model is trained using a simulation dataset to obtain the final MLP neural network model, and its surface plot is as shown in Figure 3 and the regression diagnostic plot is as shown in Figure 4 .

[0068] Step 4: Taking the minimization of the deviation of the stopping position of the accident participants or the tire friction marks in a road traffic accident as the core, construct an objective function, use the final MLP neural network model in Step 3 to predict the accident result, and obtain the optimal solution of the parameters that cannot be confirmed by a global optimization algorithm with multi-start search; use the optimal solution of the parameters that cannot be confirmed to perform simulations in the PC-Crash simulation software, compare the simulation results with the actual survey data, and verify the accuracy of the optimal solution.

[0069] When the optimal solution is accurate enough, the undetermined parameters of this accident case are optimally matched. If there is a large error between the simulation result of the optimal solution and the actual survey result, it indicates that the final MLP neural network model in Step 3 needs to be further optimized, and the MLP neural network model is further optimized by adding simulation test points in Step 2.

[0070] In this embodiment, the objective function is to minimize the absolute value deviation between the simulation result and the stopping position of the accident participants obtained from the on-site survey; taking the minimization of this objective function as the optimization goal and the parameter space of the trailer speed and angular velocity as the boundary constraint, the optimal speed and angular velocity are obtained as 91.5 km / h and 0.16 rad / s respectively.

[0071] Feed the optimal solution back to the PC-Crash simulation software for verification simulation; as shown in Figure 5 is the comparison diagram of the simulation of the optimized model and the stopping position of the vehicle at the accident scene. The average relative position error of the stopping positions of the car, trailer and semi-trailer obtained from the simulation and the survey is less than 5%, which proves the reliability of this method.

[0072] Figure 6 is the comparison diagram of the stopping position of the vehicle at the accident scene simulated by the speed and angular velocity given in Step 1; comparing Figure 5 and Figure 6 , it can be seen that this method can reduce the deviation of the stopping position of the accident participants obtained from the simulation and the survey without manual intervention, which proves that this method realizes the intelligent reproduction of road traffic accidents.

[0073] The above is only the preferred embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any modification and replacement based on the technical solutions and inventive concepts provided by the present invention should be covered by the protection scope of the present invention.

Claims

1. An intelligent road traffic accident reconstruction method based on automated Pc-Crash simulation and MLP algorithm, characterized in that, It includes the following steps: Step 1: Establish a simulation model template in Pc-Crash simulation software according to the road traffic accident scene information; Determine the road conditions at the accident site based on the general overview map of the accident scene taken at the road traffic accident site and the subsequent on-site surveying or electronic map, and establish a simulation background map or road model in Pc-Crash simulation software; Obtain the information of the accident-involved vehicles according to the road traffic accident scene investigation, and establish an accurate and reliable vehicle model in Pc-Crash simulation software; Analyze the braking marks according to the principle of traceology to obtain the mutation points of the braking mark trajectory, and then obtain the geographical location where the collision occurred; obtain the relative angle of vehicle collision by analyzing the shape of the braking marks and the deformation characteristics of the vehicle; establish a suitable relative position of the vehicles in Pc-Crash simulation software according to the geographical location where the collision occurred and the relative angle of vehicle collision; Analyze the vehicle speeds and avoidance angular velocities of the accident-involved vehicles based on video surveillance, driving recorders, and EDR devices, and establish a vehicle motion model in Pc-Crash simulation software according to the vehicle speeds and avoidance angular velocities; Step 2: Reasonably design simulation experiments in the parameter space for the parameters that cannot be confirmed at the road accident scene, and obtain a simulation data set through automated simulation using the automated simulation module according to the designed simulation experiments; The reasonable design of simulation experiments in the parameter space for the parameters that cannot be confirmed at the road accident scene is as follows: for the parameters that cannot be determined at the road accident scene, use the values given at the accident scene as reference values, and set data intervals according to the reference values; Adopt Latin hypercube sampling for high-dimensional parameter spaces and full factorial experimental design for low-dimensional parameter spaces to design experimental points; among them, high-dimensional parameters refer to data with dimensions greater than 5, and low-dimensional data refer to data with dimensions less than or equal to 5; According to the designed experimental points, simulate the accident process under each parameter through the automated simulation module, and extract the tire friction marks and the stopping positions of the accident participants, so as to form a simulation data set of the attributes and behaviors of the accident participants and the characteristics of the vehicles and marks after the accident; Step 3: Construct a multi-layer perceptron neural network model, optimize the multi-layer perceptron neural network model, and train the optimized multi-layer perceptron neural network model using the simulation data set in Step 2 to obtain the final multi-layer perceptron neural network model; Step 4: With the minimization of the deviation of the stopping position or tire friction marks of the accident participants in the road traffic accident as the core, construct an objective function, use the final multi-layer perceptron neural network model in Step 3 to predict the accident result, and obtain the optimal solution of the parameters that cannot be confirmed using a global optimization algorithm with multi-start search; perform a simulation in Pc-Crash simulation software using the optimal solution of the parameters that cannot be confirmed, and compare the simulation result with the actual survey data to verify the accuracy of the optimal solution.

2. The intelligent reproduction method of road traffic accidents based on automated Pc-Crash simulation and MLP algorithm according to claim 1, wherein The automated simulation module refers to being embedded in Pc-Crash simulation software to achieve automated simulation and output simulation results; The implementation of the automated simulation module: (1) Based on the MATLAB-Java cross-platform interface, the Java AWT Robot class API is called to simulate mouse and keyboard operations; mouse input simulation is used to perform click operations on the Pc-Crash simulation software, and keyboard input simulation is used to input specific parameters in the Pc-Crash simulation software; furthermore, the save-as operation of the accident process simulation model template and the accident speed parameter adjustment process are realized. (2) Based on the MATLAB-Java cross-platform interface, the Java AWT Robot class API and MATLAB screen capture and optical character recognition functions are called to simulate click operations, realize the simulation operation of the adjusted model, and monitor its process status by recognizing the progress bar value in the image. (3) When it is recognized that the simulation process reaches its maximum simulation time, the simulation terminates, and the output of the simulation result file is realized by simulating mouse and keyboard operations. (4) The output result is obtained by analyzing and reading the simulation result file through Matlab.

3. The intelligent road traffic accident reconstruction method based on automated Pc-Crash simulation and MLP algorithm according to claim 1, characterized in that, The multi-layer perceptron neural network model includes an input layer, a hidden layer, and an output layer; the ReLU activation function is used in the hidden layer, and the linear activation function is used in the output layer. The depth and width of the network are dynamically adjusted according to the parameter space dimension.

4. The intelligent reproduction method of road traffic accidents based on automated Pc-Crash simulation and MLP algorithm according to claim 3, characterized in that Optimizing the multi-layer perceptron neural network model means: using the Bayesian optimization algorithm to optimize the hyperparameters of the multi-layer perceptron neural network model; constructing the mapping relationship between hyperparameters and model performance through the Gaussian process model, aiming to minimize the mean square error of the validation set, and iteratively searching for the optimal hyperparameter combination; evaluating the generalization ability of the multi-layer perceptron neural network model through K-fold cross-validation.

5. The intelligent reproduction method of road traffic accidents based on automated Pc-Crash simulation and MLP algorithm according to claim 1, characterized in that, In step 4, the optimal solution is verified not to be the optimal solution by comparing the simulation result with the actual survey data. Further training the multi-layer perceptron neural network model by adding simulation test points in step 2 to improve the reliability of the multi-layer perceptron neural network model.

Images