Traffic incident digital twinborn reproduction method based on V2X data communication network

Abstract

The invention relates to a traffic incident digital twinborn reproduction method based on a V2X data communication network, and belongs to the field of intelligent traffic systems. According to the method, vehicle driving data, traffic signal lamp states and map data are collected through a vehicle-mounted unit and a road side unit, and are transmitted to a local server through a V2X network; and then, a convolution block attention module is used to extract vehicle characteristics, a bidirectional long-short-term memory network is combined to learn time sequence association between vehicles, and finally, a multi-layer perceptron is used to determine a traffic event type. Once a traffic event is detected, the system utilizes Open Street Map and Road Runner to construct a high-precision map model, SUMO simulation software is imported to reproduce the dynamic process of the traffic event, finally SUMO simulation track data is imported into a CARLA platform, a dynamic entity is generated, and details of the traffic event are reproduced.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent transportation systems and relates to a digital twin reproduction method of traffic events based on a V2X data communication network. Background Art

[0002] Current traffic incident detection and reconstruction systems usually rely on large-scale video surveillance data collection and analysis. Such systems require the deployment of a large number of cameras on roads and intersections, and collect behavioral data of vehicles, pedestrians, and other road participants through video. In this approach, traffic incident detection relies on image processing technology and computer vision algorithms, which usually involve complex processes such as object recognition, trajectory tracking, and behavior prediction. However, these systems have several significant limitations.

[0003] First, the processing and analysis of video surveillance data requires a lot of human and material resources. For example, in order to accurately extract traffic behavior from the video, professionals are usually required to label and verify the video data. Secondly, the storage and transmission costs of video data are extremely high, especially when long-term monitoring and storage are required, the storage device and bandwidth requirements will rise rapidly. In addition, video data analysis also requires high-performance computing resources to process large-scale video streams in real time, which increases the financial consumption of the system.

[0004] In terms of the generation of structured data, the processing results of video data usually contain a lot of unnecessary information, such as background environment, noise information, etc., which leads to a lot of redundancy in the structured data generated, which is not conducive to the subsequent reconstruction of digital twin scenes. A large amount of redundant information will increase the complexity of reproducing traffic events and is not conducive to the accurate reproduction of events. Especially in digital twin systems, this redundant information may affect the true restoration of events, causing key details to be concealed, and weakening the efficiency and accuracy of the system.

[0005] In contrast, the traffic incident detection and replay method based on V2X (Vehicle-to-Everything) data can effectively avoid these problems. By directly collecting structured data from vehicles and roadside equipment, it can accurately record vehicle status and environmental information and reduce the generation of redundant data. At the same time, this method avoids the high cost of video systems, lowers the threshold for traffic incident detection and replay, and improves the efficiency and accuracy of the system. Summary of the invention

[0006] In view of this, the purpose of the present invention is to provide a traffic event digital twin reproduction method based on the V2X data communication network to solve the problems of inaccurate traffic event detection and high difficulty in reproducing traffic accidents. The system consists of two parts: a module for detecting traffic events through V2X data and a traffic accident reproduction module.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] A traffic event digital twin reproducing method based on a V2X data communication network, the method comprising the following steps:

[0009] S1: The OBUs at the intersection where the accident occurred collect driving data of the vehicle in real time through the vehicle Ethernet. The OBU terminal establishes a vehicle-to-everything (V2X) communication connection with the RSU within the coverage of the roadside RSU signal, collects traffic light and map-related data, and transmits the acquired V2X data to the local server through the RSU device;

[0010] S2: In the local server, the obtained V2X data is pre-processed;

[0011] S3: Input the processed data of each vehicle into the convolutional block attention module to learn the important features of each vehicle;

[0012] S4: The feature-extracted data is then fed into a bidirectional long short-term memory network to learn the global temporal association between each vehicle.

[0013] S5: The learned data is passed through the multi-layer perceptron module to determine what kind of traffic event it belongs to;

[0014] S6: Based on the traffic event that is judged to have occurred, a digital twin is used to generate scenes and dynamic objects in the simulator, and the traffic event is restored based on previous data.

[0015] Furthermore, in S1, the V2X data includes parameters of the vehicle's position, speed, acceleration, direction, throttle status, braking status and steering angle, which fully reflect the vehicle's driving status and driving behavior; through the V2X connection, the OBU can receive and integrate status information from traffic lights, surrounding road conditions and map data in real time, ensuring that driving data and key information of the traffic environment are updated synchronously.

[0016] Furthermore, in S2, the obtained V2X raw data is decoded, and the data of each vehicle is classified in chronological order; V2X data consists of three types: basic safety message BSM, signal light status and timing SPaT, and map data MAP. Each type of data is closely coordinated in time and space dimensions to jointly realize the detection and reproduction of traffic events; BSM data describes the dynamic state of the vehicle and is the core data source reflecting the vehicle behavior, including timestamp, latitude and longitude and altitude, speed, acceleration, direction angle, azimuth velocity, braking status, and information on historical path points; BSM also records The size information of the vehicle provides a geometric basis for the accurate modeling of traffic events; SPaT data is used to reflect the dynamic status of traffic lights, including intersection ID and signal group ID, current status of the traffic light, minimum and maximum green light duration; through timestamps, SPaT data is associated with vehicle status data in BSM to support in-depth analysis of the interaction between vehicles and traffic lights; MAP data provides a geographical basis for the spatial modeling of traffic events, including reference points for setting map reference coordinates, point sets for defining road geometry, lane IDs for matching the specific lanes where vehicles are located, and continuous lane indexes.

[0017] Further, in S3, in the convolutional block attention module, in the channel attention module, for a given feature map F∈R C×H×W ; First perform average pooling and maximum pooling to obtain two feature maps: and The resulting formula is as follows:

[0018]

[0019] Where H and W are the length and width of the input image respectively. is the feature vector after average pooling, To perform the maximum pooling of the feature vector, F(i, j) represents the calculation of the number in the i-th row and j-th column;

[0020] The pooled feature map is then input into a shared fully connected layer to generate the channel attention weight, the formula is as follows:

[0021]

[0022] Among them, σ is the Sigmoid activation function, M c Channel attention weight, FC means full connection layer calculation; then the attention weight M c Applied to the original feature image, we get the weighted output, the formula is as follows:

[0023] F'=M c ·F

[0024] Where F is the original feature map, and F is the feature map obtained after calculation;

[0025] Given an input feature F∈R in the temporal attention module T×C , first perform average pooling and maximum pooling along the channel dimension, the formula is as follows:

[0026]

[0027] Among them, C is the number of channels of the input feature, and F(t, k) represents the calculation of the t-th feature channel. and They are two vectors of length T, representing the average and maximum feature values ​​in the time dimension;

[0028] Then you will get and Concatenate to form a vector of size T×2, and use a convolution kernel to convolve the concatenated features to generate the time attention weight. The formula is as follows:

[0029]

[0030] Among them, σ is the Sigmoid activation function, M t is the time attention weight, Conv means convolution calculation; then the time attention weight is applied to the features of each time step to obtain the time weighted graph, the formula is as follows:

[0031] F'=M t ·F

[0032] Where F is the original feature map, and F is the feature map obtained after calculation.

[0033] Further, in S4, for the bidirectional long short-term memory network, the given input sequence X=(x 1 , x 2 , ..., x t ), where x t is the input of time step t, and the bidirectional long short-term memory network calculates the forward hidden state and backward hidden state of each time step and combines them together;

[0034] The calculation formula for the hidden state in the forward long short-term memory network is as follows:

[0035]

[0036] in, Represents the hidden state of the forward long short-term memory network at time t-1, LSTM forwardIndicates that the forward calculation of LSTM is performed. represents the hidden state of the forward long short-term memory network at time t;

[0037] The calculation formula for the hidden state in the backward long short-term memory network is as follows:

[0038]

[0039] in, Represents the hidden state of the forward long short-term memory network at time t+1, LSTM backward Indicates that the backward calculation of LSTM is performed. represents the hidden state of the backward long short-term memory network at time t;

[0040] The forward and backward hidden states are concatenated together at the output layer as the final hidden state of the bidirectional long short-term memory network at time t:

[0041]

[0042] Among them, h t is the hidden state of the bidirectional long short-term memory network at time t, containing the contextual information obtained from the algorithm;

[0043] The internal calculations in the bidirectional long short-term memory network include input gate, forget gate, output gate, candidate memory gate, update memory state and update hidden state. The specific formula is as follows:

[0044]

[0045] The above formula is the calculation formula for input gate, forget gate, output gate, candidate memory gate, update memory state and update hidden state, where σ, x t , W x , W h 、h t-1 , b, and tanh represent the Sigmoid activation function, the input matrix, the weight matrix of the current gate and the input matrix, the weight matrix of the current gate and the hidden state of the previous time step, the hidden state of the previous time step, the bias term of the current gate, and the Tanh activation function, respectively; i t 、f t , o t , ct and ht represent the calculation result after the input gate, the calculation result after the forget gate, the calculation result after the output gate, the calculation result after the candidate memory gate, the calculation result after the update memory state algorithm, and the calculation result after the update hidden state algorithm, respectively.

[0046] Furthermore, in S5, the multilayer perceptron package includes 3 hidden layers and one output layer, and the calculation formula is as follows:

[0047]

[0048] Among them, h is the hidden layer output, f uses the ReLU activation function, and W 1，2 represents the weight matrix, b 1，2 represents the bias value; y is the final output, and g uses the softmax activation function for multi-classification;

[0049] The output is a 3D vector, which represents the probability of three traffic events: vehicle accidents, traffic congestion and failure to comply with traffic regulations.

[0050] Further, in S6, after a vehicle traffic accident is detected, time information is used to backtrack, extract information from the original data, and restore the traffic event in the simulator; first, after the system detects the specific time, location, and related entities of the accident, the system backtracks forward or backward according to the timestamp information to extract information before and after the accident from the recorded original data; the information includes the vehicle speed, acceleration, lane position, and motion state at the time of the accident; the extracted information is organized and structured, and imported into the simulator as input data;

[0051] In the process of reproducing the traffic event scene, Open Street Map is first used to obtain the basic map data of the target area, including the geographic information of the road network, lane information, traffic signs and signal light locations; this data is then imported into the Road Runner tool for editing and optimization, and traffic infrastructure information is added manually or automatically, including accurate lane distribution, lane width, signal light logic, crosswalks and the location of static objects on the roadside, to finally generate a high-precision map model that can support simulation; next, the generated map is imported into the SUMO simulation software, and the dynamic information extracted from the V2X data, namely the location, speed, acceleration and signal light status changes of vehicles and vehicles, is used to set traffic flow parameters and simulation configurations, reproduce the dynamic process of traffic events in SUMO, and generate simulated vehicle and interaction behavior data; then, the simulation trajectory data of SUMO is imported into the CARLA simulation platform, the high-precision map environment is loaded synchronously, and dynamic entities are generated in the scene based on the trajectory data; through the physical simulation of CARLA, the details of the relative movement and interaction behavior of vehicles in the traffic event are reproduced.

[0052] The beneficial effects of the present invention are as follows: the method realizes data collection based on V2X communication through an onboard unit (OBU) and a roadside unit (RSU), and directly obtains high-precision target information from vehicles and traffic infrastructure. The OBU collects dynamic information such as vehicle speed, acceleration, direction angle, position, etc., and combines the signal light state SPaT (such as current state, minimum and maximum green light duration) and map scene data MAP (such as road geometry, lane ID and continuous lane index) broadcast by the RSU to realize efficient collection of multi-source traffic data at a lower cost and higher real-time performance. Unlike video stream processing, V2X communication directly provides structured data without the need for complex target recognition and information extraction processes, greatly reducing system resource consumption, and avoiding data loss or errors caused by environmental interference. The collected V2X data is decoded and preprocessed to extract vehicle, signal light and map information, and local features are extracted through a convolutional neural network (CNN) combined with an attention mechanism to comprehensively analyze vehicle dynamics and traffic environment. Subsequently, a bidirectional long short-term memory network (BiLSTM) is used to model the temporal dependency between vehicles, signal lights and road scenes. Finally, the type of traffic incidents was identified through a multi-layer perceptron (MLP) classification model, and an accurate road network model was built in combination with Open Street Map and Road Runner. The V2X data was input into the SUMO simulation software to restore and dynamically reproduce the event scene. This method fully utilizes the accuracy, structure and real-time advantages of V2X broadcast data, which not only greatly improves the efficiency and accuracy of traffic incident detection and reproduction, but also avoids the high cost and limitations of traditional video stream processing, providing an efficient and reliable solution for intelligent transportation and accident analysis.

[0053] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0055] Figure 1 It is a system structure diagram of the present invention;

[0056] Figure 2 The present invention is a framework for traffic event detection and traffic event reproduction;

[0057] Figure 3 Reproduce the flow chart for the traffic incident. DETAILED DESCRIPTION

[0058] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0059] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0060] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0061] V2X data consists of three types: BSM (Basic Safety Message), SPaT (Signal State and Timing), and MAP (Map Data). These types of data work closely together in time and space to detect and reproduce traffic events. Table 1 shows the V2X data used.

[0062] Table 1

[0063]

[0064]

[0065] BSM data is the core data source for describing the dynamic state of vehicles, and is mainly used to reflect the real-time behavior and state of vehicles in traffic events. It contains multiple key fields, such as timestamp, which is used to record the specific time when data is generated to ensure the alignment of events in the time dimension; longitude, latitude and altitude information, which accurately identifies the current position of the vehicle; speed and acceleration fields describe the vehicle's motion state, such as sudden braking or rapid acceleration; direction angle and azimuth velocity are used to characterize the vehicle's driving direction and rotation information, which is particularly suitable for analyzing the vehicle's turning or lane changing behavior. In addition, BSM also contains the vehicle's braking status information, which is used to determine whether emergency braking has occurred, and historical trajectory points, which are used to record the vehicle's past motion trajectory, which is convenient for subsequent trajectory backtracking and analysis. The vehicle's dimensional data (such as length, width and height) provides a physical basis for the geometric modeling of traffic events. This information is not only used for the dynamic reproduction of the vehicle, but also provides data support for the interactive analysis with other traffic elements.

[0066] SPaT data is used to describe the dynamic state of traffic lights and their changing patterns, providing key information for signal control analysis in traffic event detection and reproduction. SPaT data contains intersection ID and signal group ID, which clearly identify the specific location and grouping of the traffic lights, facilitating spatial association with the vehicle status in the BSM data; the current state field of the traffic light records the real-time state of the traffic light, such as the changes in the red, green or yellow light, which is an important basis for determining whether a vehicle has violated the rules; the minimum and maximum green light time fields predict the duration range of the traffic light in the current state, providing data support for traffic signal optimization analysis. By matching with the timestamps in the BSM data, SPaT data can help analyze the dynamic interaction of vehicles with traffic lights within a specific time, especially in traffic events such as running red lights and sudden braking.

[0067] MAP data is used to provide the geographical and road structure basis of traffic event scenes, mainly describing the geometry and topological relationship of the road network. MAP data includes reference point fields, which serve as the base coordinates of the map and are used to locate all road elements; point set fields define the geometry of the road and describe the connection relationship between multiple segments of coordinate points, providing a basis for high-precision road modeling; lane IDs and continuous lane indexes are used to identify specific roads and their continuity, making it easier to determine the spatial relationship between vehicles and roads. In the re-enactment of traffic events, MAP data is combined with the vehicle position information of BSM and the signal light position of SPaT to achieve accurate positioning of vehicles in the road environment, while supporting the re-enactment of road network characteristics (such as complex scenes such as multiple lanes and intersections). This information provides a complete geographical background and scenario support for the digital twin re-enactment of traffic events.

[0068] Figure 1The overall structure framework of the system is mainly divided into three modules: data collection, event detection and event reproduction. The overall description is as follows:

[0069] In the data collection module, the vehicle's own data (such as speed, position, acceleration, driving direction, etc.) and the surrounding environment data (such as traffic light status, etc.) monitored by the roadside equipment (RSU) are collected. All of this data is transmitted to the local server in real time through the forwarding equipment of the vehicle Ethernet and V2X network for integration and analysis. Through the real-time collection and integration of this multi-source data, the system can obtain comprehensive traffic information to ensure the accuracy of event detection and the integrity of scene restoration.

[0070] Before entering the event detection module, the received data will first go through a rigorous data preprocessing step. This step cleans, denoises, and converts the raw data to ensure data quality and consistency. Next, the system extracts various types of information by category, including key driving data such as vehicle speed, position, acceleration, the status and cycle of traffic lights, and map information. In this way, the structured data generated by the preprocessing module provides accurate and redundant input for the event detection algorithm, reducing data noise during the calculation process.

[0071] In the event detection module, the processed data is input into the event detection algorithm based on machine learning or rules for calculation. The algorithm determines whether there is a potential traffic incident based on the relative position, driving direction, speed and other characteristics of the vehicle. Once a possible traffic incident is detected, the system will extract the key information of the incident, including the specific data of the vehicle involved, the geographical location of the incident, and related spatiotemporal information. These structured event data can provide reliable support for subsequent scene reproduction.

[0072] In the traffic incident re-enactment module, the system first uses the open source map resource Open Street Map (OSM) and the traffic simulation modeling tool Road Runner to perform high-precision map modeling of the area where the incident occurred. Through Open Street Map, the system can obtain detailed geographic information such as road structure, traffic signs, and surrounding buildings, while Road Runner provides map customization and fine-tuning functions to ensure that the modeling conforms to the actual traffic environment. The map model generated at this stage will provide a realistic and accurate three-dimensional scene for the simulation of traffic incidents.

[0073] After completing the map modeling, the system imports the map data and the extracted traffic event data into the traffic simulation platform SUMO (Simulation of Urban MObility). In the SUMO simulation environment, the system reproduces the occurrence process of traffic events, reproducing the vehicle's movement trajectory, collision scenarios, traffic light changes and other details.

[0074] Figure 2 This is the system traffic incident detection model architecture, which is described as follows:

[0075] When training the model, the system first pre-processes the collected V2X data. This data includes status information such as vehicle speed, acceleration, location, and driving direction, as well as key environmental information such as map geographic data and traffic light status. Since these data sources come from different vehicles and roadside equipment, the raw data must first be decoded to ensure that each data item can be correctly parsed and various key information can be extracted from it to prepare for subsequent data processing and feature extraction.

[0076] Next, the decoded data is processed for dimensionality reduction to reduce data redundancy and feature complexity. The system performs One-Hot encoding on the categorical feature of the traffic light status and converts the character features of the traffic light into integer values. This not only retains the feature differences of different traffic light states, but also avoids the misunderstanding that the categorical data has a sequential relationship in terms of values. In addition, for missing data in the data set, the system will use appropriate filling methods based on business needs and data characteristics to ensure data integrity and consistency. In the last step of data preprocessing, the system uses data normalization to map all feature data to the range of 0 to 1, which not only helps to reduce the differences between data, but also accelerates the convergence of the model and improves training efficiency and stability.

[0077] In the feature extraction stage, the system constructs a convolutional block-attention module to extract local features of V2X data. In this module, each vehicle data will be combined with the corresponding map and traffic light data, and the local perception mechanism of the convolutional layer will be used to extract key spatial features, and the weights of important features will be enhanced through the attention mechanism. The reason for using the Sigmoid activation function here is that the Sigmoid function can compress the output between 0 and 1, which is suitable for emphasizing or weakening the weight of features in the attention mechanism. The smooth curve of Sigmoid enables the model to gently highlight or suppress each feature, thereby obtaining a richer feature expression. In the convolutional block attention module, for a given feature map F∈R in the channel attention module C×H×W ; First perform average pooling and maximum pooling to obtain two feature maps: and The resulting formula is as follows:

[0078]

[0079] The pooled feature map is then input into a shared fully connected layer to generate the channel attention weight, the formula is as follows:

[0080]

[0081] Among them, σ is the Sigmoid activation function, M c Channel attention weight. Then the attention weight M c Applied to the original feature image, we get the weighted output, the formula is as follows:

[0082] F'=M c ·F

[0083] Given an input feature F∈R in the temporal attention module T×C , first perform average pooling and maximum pooling along the channel dimension, the formula is as follows:

[0084]

[0085] Among them, the obtained and They are two vectors of length T, representing the average and maximum feature values ​​in the time dimension.

[0086] Then you will get and Concatenate to form a vector of size T×2, and use a convolution kernel to convolve the concatenated features to generate the time attention weight. The formula is as follows:

[0087]

[0088] Among them, σ is the Sigmoid activation function, M t is the temporal attention weight. Then the temporal attention weight is applied to the features of each time step to obtain the temporal weighted graph, and the formula is as follows:

[0089] F'=M t ·F

[0090] Based on the extracted local features, the system inputs the data into a bidirectional long short-term memory network (BiLSTM) to learn the temporal relationship between vehicles after combining map and traffic light data. Bidirectional LSTM can capture complex dependencies in the time dimension, enabling the model to focus on both past and future states, thereby more accurately describing the interaction between vehicles in traffic events.

[0091] When training the BiLSTM model, in order to balance model performance and computing resources, the following carefully designed hyperparameter configurations were used: First, the number of neurons in the hidden layer of the model was set to 128. This dimension can balance the richness of feature expression and computing efficiency, ensuring that the complex features in the input data can be effectively captured. In order to further improve the model's temporal learning ability, the network is designed as a bidirectional structure with two layers of LSTM stacked. This stacked design enables the model to extract information from deeper time steps and adapt to dependencies with larger time spans.

[0092] In addition, we set the time step length to 50, which can capture medium-length time dependencies and have sufficient perception of longer patterns in time series. At the same time, the learning rate is set to 0.001, and the Adam optimizer is used. Adam can adaptively adjust the learning rate, update parameters in a more robust way, and speed up the convergence process. The batch size is set to 128, which is suitable for larger-scale training data, ensuring that each update can reflect the characteristics of the overall data, while improving the utilization of parallel resources such as GPUs.

[0093] In terms of regularization, a Dropout ratio of 0.2 is used to prevent overfitting and ensure that the model maintains good generalization performance when facing new data. In order to avoid the gradient explosion problem, the gradient clipping threshold of the model is set to 5 to ensure that the gradient will not be too large during the back propagation process and maintain a stable training process. At the same time, the Xavier initialization method is selected for the weight initialization of the model, which makes the distribution of the initial weights more reasonable, thereby accelerating the early convergence of the network and improving the training efficiency.

[0094] For the given input sequence X=(x 1 ,x 2 ,…,x t ), where x t is the input of time step t, and the bidirectional long short-term memory network calculates the forward hidden state and backward hidden state of each time step and combines them together.

[0095] The calculation formula for the hidden state in the forward long short-term memory network is as follows:

[0096]

[0097] in represents the hidden state of the forward LSTM network at time t.

[0098] The calculation formula for the hidden state in the backward long short-term memory network is as follows:

[0099]

[0100] in represents the hidden state of the backward LSTM network at time t.

[0101] The forward and backward hidden states are concatenated together at the output layer as the final hidden state of the bidirectional long short-term memory network at time t:

[0102]

[0103] Among them, h t is the hidden state of the bidirectional LSTM network at time t, which contains the context information obtained from the algorithm.

[0104] The internal calculation in the bidirectional long short-term memory network includes the following parts: input gate, forget gate, output gate, candidate memory gate, update memory state and update hidden state. The specific formula is as follows:

[0105]

[0106] The above formulas are the calculation formulas for input gate, forget gate, output gate, candidate memory gate, update memory state and update hidden state from top to bottom, where σ, x t , W x , W h 、h t-1 , b, and tanh represent the Sigmoid activation function, the input matrix, the weight matrix of the current gate and the input matrix, the weight matrix of the current gate and the hidden state of the previous time step, the hidden state of the previous time step, the bias term of the current gate, and the Tanh activation function, respectively.

[0107] Here, Sigmoid and Tanh activation functions are used in BiLSTM: the Tanh function is suitable for mapping data to the interval of -1 to 1, which helps to capture the characteristics of positive and negative changes; while the output range of the Sigmoid function is between 0 and 1, which is suitable for the "gating" operation in the model to control the flow of information in the time series, thereby optimizing the expression effect of time series data. The combination of the two enables BiLSTM to better learn complex nonlinear relationships in the time dimension and accurately grasp the important time series characteristics in traffic events.

[0108] After the BiLSTM module completes processing, the system inputs the learned high-dimensional features into the multi-layer perceptron (MLP) for final traffic event type identification.

[0109] The multilayer perceptron package contains 3 hidden layers and one output layer. The calculation formula is as follows:

[0110]

[0111] Among them, h is the hidden layer output, f uses the ReLU activation function, W represents the weight matrix, and b represents the bias value; y is the final output, and g uses the softmax activation function for multi-classification.

[0112] 128 neurons are used in the hidden layer, and the output is a 3D vector, which represents the probability of three traffic events: vehicle accidents, traffic congestion, and failure to comply with traffic regulations.

[0113] MLP maps multidimensional features to the output space through a fully connected layer to determine whether a specific traffic event occurred during the time period and the type of event. The final judgment result will be returned to the original data for time search to find the time period of the relevant event and the vehicle data involved.

[0114] After completing the event recognition, the system extracts the vehicles within the time period when the event occurred and imports this data into the traffic event restoration module to achieve accurate reproduction of the actual traffic event.

[0115] In the process of reproducing the traffic event scene, Open Street Map is first used to obtain the basic map data of the target area, including geographic information such as road network, lane information, traffic signs and signal light locations. Then this data is imported into the Road Runner tool for editing and optimization, and detailed traffic infrastructure information such as accurate lane distribution, lane width, signal light logic, crosswalks and the location of static objects on the roadside are added manually or automatically, and finally a high-precision map model that can support simulation is generated. Next, the generated map is imported into the SUMO simulation software, and the dynamic information extracted from the V2X data (such as the location, speed, acceleration and signal light status changes of vehicles and vehicles) is used to set traffic flow parameters and simulation configurations, reproduce the dynamic process of traffic events in SUMO, and generate simulated vehicle and interaction behavior data. Subsequently, the simulation trajectory data of SUMO is imported into the CARLA simulation platform, the high-precision map environment is loaded synchronously, and dynamic entities (including vehicles) are generated in the scene based on the trajectory data. Through CARLA's high-fidelity physical simulation, the details of the relative movement and interactive behavior of vehicles in traffic events are truly reproduced.

[0116] Figure 3 The flow chart for reproducing traffic events is as follows:

[0117] First, at time t 1 , the system receives a MAP message (map scene data), which contains static scene data such as the geometry of the road, lane identification information, and continuous lane indexes. These data are used to build an accurate road network model as the basis for reproducing the entire scene.

[0118] Then, from time point t2 Initially, the system receives a series of BSM messages (Basic Safety Messages). BSM messages contain dynamic information about the vehicle, such as speed, acceleration, direction angle, and position. At each time step t 2 , t 3 , t 4 , t 5 The BSM data is updated in real time to describe the state changes of the vehicle in the scene. By gradually superimposing the BSM data of each time step, the system can accurately restore the trajectory of the vehicle in the scene and its dynamic behavior.

[0119] Figure 2 As shown in Figure 2, at each time step, the dynamic data is superimposed on the static scene data (provided by MAP), thereby realizing the dynamic evolution of the vehicle in the real road environment. For example, at time point t 2 , a small number of vehicles enter the scene; as time advances to t 3 and t 4 , the number and position of vehicles change, simulating the change process of traffic flow; finally at t 5 , some vehicles approached or collided.

[0120] In this way, the system uses V2X data as the core driving force, combines static road network scenes with dynamic traffic behaviors, and reproduces the process of traffic accidents step by step. Compared with the traditional method that relies on video streams, this method has higher data accuracy and real-time performance, and can also describe the spatiotemporal evolution characteristics of traffic events in detail, providing strong support for traffic accident analysis and research.

[0121] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A traffic event digital twin reproduction method based on a V2X data communication network, characterized in that: The method comprises the following steps: S1: The OBUs at the intersection where the accident occurred collect driving data of the vehicle in real time through the vehicle Ethernet. The OBU terminal establishes a vehicle-to-everything (V2X) communication connection with the RSU within the coverage of the roadside RSU signal, collects traffic light and map-related data, and transmits the acquired V2X data to the local server through the RSU device; S2: In the local server, the obtained V2X data is pre-processed; S3: Input the processed data of each vehicle into the convolutional block attention module to learn the important features of each vehicle; S4: The feature-extracted data is then fed into a bidirectional long short-term memory network to learn the global temporal association between each vehicle. S5: The learned data is passed through the multi-layer perceptron module to determine what kind of traffic event it belongs to; S6: Based on the traffic event that is judged to have occurred, a digital twin is used to generate scenes and dynamic objects in the simulator, and the traffic event is restored based on previous data.

2. The method for reproducing a traffic event digital twin based on a V2X data communication network according to claim 1, characterized in that: In S1, the V2X data includes parameters of the vehicle's position, speed, acceleration, direction, throttle status, braking status and steering angle, which fully reflect the vehicle's driving status and driving behavior; through the V2X connection, the OBU can receive and integrate status information from traffic lights, surrounding road conditions and map data in real time, ensuring that driving data and key information of the traffic environment are updated synchronously.

3. The method for reproducing a traffic event digital twin based on a V2X data communication network according to claim 1, characterized in that: In the S2, the obtained V2X raw data is decoded, and the data of each vehicle is classified in chronological order; V2X data consists of three types: basic safety message BSM, signal light status and timing SPaT, and map data MAP. Various types of data are closely coordinated in time and space dimensions to jointly realize the detection and reproduction of traffic events; BSM data describes the dynamic state of the vehicle and is the core data source reflecting the vehicle behavior, including timestamp, longitude and latitude and altitude, speed, acceleration, direction angle, azimuth velocity, braking status, and information of historical path points; BSM also records the size information of the vehicle to provide a geometric basis for accurate modeling of traffic events; SPaT data is used to reflect the dynamic state of traffic lights, including intersection ID and signal group ID, current state of the light, minimum and maximum green light duration; Through timestamps, SPaT data is associated with vehicle status data in BSM, supporting in-depth analysis of the interaction between vehicles and traffic lights; MAP data provides a geographical basis for spatial modeling of traffic events, including reference points for setting map reference coordinates, point sets for defining road geometry, lane IDs for matching the specific lanes where vehicles are located, and continuous lane indexes.

4. The method for reproducing a traffic event digital twin based on a V2X data communication network according to claim 1, characterized in that: In S3, in the convolutional block attention module, in the channel attention module, for a given feature map F∈R C×H×W ; First perform average pooling and maximum pooling to obtain two feature maps: and The resulting formula is as follows: Where H and W are the length and width of the input image respectively. is the feature vector after average pooling, is the feature vector after the maximum pooling, F(i, j) represents the calculation of the number in the i-th row and j-th column; The pooled feature map is then input into a shared fully connected layer to generate the channel attention weight, the formula is as follows: Among them, σ is the Sigmoid activation function, M c Channel attention weight, FC means full connection layer calculation; then the attention weight M c Applied to the original feature image, we get the weighted output, the formula is as follows: F’=M c ·F Where F is the original feature map, and F is the feature map obtained after calculation; Given an input feature F∈R in the temporal attention module T×C , first perform average pooling and maximum pooling along the channel dimension, the formula is as follows: Among them, C is the number of channels of the input feature, and F(t, k) represents the calculation of the t-th feature channel. and They are two vectors of length T, representing the average and maximum feature values ​​in the time dimension; Then you will get and Concatenate to form a vector of size T×2, and use a convolution kernel to convolve the concatenated features to generate the time attention weight. The formula is as follows: Among them, σ is the Sigmoid activation function, M t is the time attention weight, Conv means convolution calculation; then the time attention weight is applied to the features of each time step to obtain the time weighted graph, the formula is as follows: F’=M t ·F Where F is the original feature map, and F is the feature map obtained after calculation.

5. The method for reproducing a traffic event digital twin based on a V2X data communication network according to claim 1, characterized in that: In S4, for the bidirectional long short-term memory network, the given input sequence X=(x1, x2, ..., x t ), where x t is the input of time step t, and the bidirectional long short-term memory network calculates the forward hidden state and backward hidden state of each time step and combines them together; The calculation formula for the hidden state in the forward long short-term memory network is as follows: in, Represents the hidden state of the forward long short-term memory network at time t-1, LSTM forward Indicates that the forward calculation of LSTM is performed. represents the hidden state of the forward long short-term memory network at time t; The calculation formula for the hidden state in the backward long short-term memory network is as follows: in, Represents the hidden state of the forward long short-term memory network at time t+1, LSTM backward Indicates that the backward calculation of LSTM is performed. represents the hidden state of the backward long short-term memory network at time t; The forward and backward hidden states are concatenated together at the output layer as the final hidden state of the bidirectional long short-term memory network at time t: Among them, h t is the hidden state of the bidirectional long short-term memory network at time t, containing the contextual information obtained from the algorithm; The internal calculations in the bidirectional long short-term memory network include input gate, forget gate, output gate, candidate memory gate, update memory state and update hidden state. The specific formula is as follows: The above formula is the calculation formula for input gate, forget gate, output gate, candidate memory gate, update memory state and update hidden state, where σ, x t , W x , W h 、h t-1 , b, and tanh represent the Sigmoid activation function, the input matrix, the weight matrix of the current gate and the input matrix, the weight matrix of the current gate and the hidden state of the previous time step, the hidden state of the previous time step, the bias term of the current gate, and the Tanh activation function, respectively; i t 、f t , o t , c t and h t They respectively represent the calculation result after passing through the input gate, the calculation result after passing through the forget gate, the calculation result after passing through the output gate, the calculation result after passing through the candidate memory gate, the calculation result after the memory state update algorithm, and the calculation result after the hidden state update algorithm.

6. The method for reproducing a traffic event digital twin based on a V2X data communication network according to claim 1, characterized in that: In S5, the multilayer perceptron package includes 3 hidden layers and one output layer, and the calculation formula is as follows: Among them, h is the hidden layer output, f uses the ReLU activation function, and W 1，2 represents the weight matrix, b 1，2 represents the bias value; y is the final output, and g uses the softmax activation function for multi-classification; The output is a 3D vector, which represents the probability of three traffic events: vehicle accidents, traffic congestion and failure to comply with traffic regulations.

7. The method for reproducing a traffic event digital twin based on a V2X data communication network according to claim 1, characterized in that: In S6, after a vehicle traffic accident is detected, time information is used to backtrack, extract information from the original data, and restore the traffic event in the simulator; first, after the system detects the specific time, location, and related entities of the accident, the system backtracks forward or backward according to the timestamp information to extract information before and after the accident from the recorded original data; the information includes the vehicle speed, acceleration, lane position, and motion state at the time of the accident; the extracted information is organized and structured, and imported into the simulator as input data; In the process of recreating traffic incident scenes, we first use Open Street Map to obtain basic map data of the target area, including geographic information of the road network, lane information, and the location of traffic signs and signal lights; This data is then imported into the Road Runner tool for editing and optimization. Traffic infrastructure information is added manually or automatically, including precise lane distribution, lane width, signal light logic, crosswalks, and the location of static objects on the roadside, to ultimately generate a high-precision map model that can support simulation. Next, the generated map is imported into the SUMO simulation software, and the dynamic information extracted from the V2X data, namely the location, speed, acceleration, and signal light status changes of vehicles and vehicles, is used to set traffic flow parameters and simulation configurations, reproduce the dynamic process of traffic events in SUMO, and generate simulated vehicle and interaction behavior data. Subsequently, SUMO's simulation trajectory data is imported into the CARLA simulation platform, the high-precision map environment is loaded synchronously, and dynamic entities are generated in the scene based on the trajectory data; through CARLA's physical simulation, the details of the relative motion and interactive behavior of vehicles in traffic events are reproduced.

Images