A stochastic traffic flow simulation method based on graph space-time neural network

Abstract

The application relates to a random vehicle flow simulation method based on a graph space-time neural network, which comprises the following steps: S1: correcting, detecting and tracking vehicles to form effective trajectory data of vehicle flow by using bridge surface videos collected by a camera; S2: establishing a vehicle departure model based on a multilayer Monte Carlo model and a graph space-time neural network; and S3: establishing a vehicle driving micro-simulation model based on the graph space-time neural network model. Compared with the prior art, the application realizes a vehicle flow simulation method conforming to actual speed change and lane change behaviors, and simultaneously establishes a vehicle simulation model of vehicle speed change-lane change coupling; the random vehicle flow simulation method based on the graph space-time neural network has strong portability and important reference significance for vehicle trajectory course analysis of bridges and other roads.

Description

Technical Field

[0001] The present invention relates to a random traffic flow simulation method, in particular to a random traffic flow simulation method based on a graph spatiotemporal neural network. Background Art

[0002] With the acceleration of urbanization, the continuous increase in traffic volume has put tremendous pressure on infrastructure such as bridges. Bridge design and maintenance require accurate simulation of vehicle loads to assess their fatigue reliability.

[0003] Traditional traffic flow simulation methods, such as models based on cellular automata, can simulate the basic dynamics of traffic flow, but often have difficulty accurately capturing the microscopic behavior of vehicles, such as speed and lane changes.

[0004] In recent years, deep learning, especially graph neural networks, has shown great potential in traffic flow prediction. Models such as the Spatio-Temporal Graph Neural Network (STGCN) can leverage graph structures to simulate the complex interactions in traffic networks, improving prediction accuracy.

[0005] Although existing deep learning models have made some progress in traffic prediction, how to apply these models to actual traffic flow simulation, especially in specific scenarios such as bridges, remains a challenge. Summary of the Invention

[0006] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a random traffic flow simulation method based on a graph spatiotemporal neural network, realize a traffic flow simulation method that conforms to actual speed and lane change behaviors, and at the same time establish a vehicle simulation model with vehicle speed change and lane change coupling; the proposed random traffic flow simulation method based on a graph spatiotemporal neural network has strong portability and important reference significance for vehicle trajectory history analysis on bridges and other roads.

[0007] The purpose of the present invention can be achieved by the following technical solutions:

[0008] The present invention provides a random traffic flow simulation method based on a graph spatiotemporal neural network, comprising the following steps:

[0009] S1: Bridge surface videos captured by cameras are corrected, detected, and tracked to generate effective trajectory data for traffic.

[0010] S2: Using the effective trajectory data of the traffic flow formed in S1, a vehicle departure model for entering the bridge is established based on the multi-layer Monte Carlo model and the graph spatiotemporal neural network;

[0011] S3: Using the effective trajectory data of the traffic flow formed in S1 and the vehicle departure model on the bridge established in S2, a vehicle driving micro-simulation model is established based on the graph spatiotemporal neural network model to realize random traffic flow simulation.

[0012] Furthermore, S1 specifically includes the following steps:

[0013] Video rectification: Correct image distortion through photogrammetric projection transformation using ground control points and a global coordinate station;

[0014] Vehicle detection and tracking: The corrected video is processed based on target detection and tracking algorithms, recording the current timestamp, vehicle number, geometric center position, and vehicle type, and calculating vehicle speed and acceleration. The acquired data together form effective trajectory data for the vehicle flow.

[0015] Furthermore, the video correction process uses ground control points and a total coordinate station to perform photogrammetric projection transformation to correct image distortion and convert the original image coordinates into image coordinates of physical space relationships, specifically including:

[0016] The original image is distorted using the photogrammetry method using Ground Control Points (GCPs) and the total coordinate station;

[0017] in:

[0018] Collecting Ground Control Points (GCPs): First, control points need to be collected on the ground at known geographical locations. These points can be clearly identified in the image and their precise coordinates are obtained using ground measurement equipment (such as GPS).

[0019] Distortion Correction: These control points are then used to correct image distortion using a photogrammetric projective transformation. This typically involves using specialized software such as ENVI or ArcGIS Pro, which can calculate the image's distortion parameters based on the known coordinates of the control points and their positions in the image, and apply the appropriate correction algorithm.

[0020] Coordinate transformation: After distortion correction, the original image coordinates are transformed into image coordinates related to physical space. This process may include mapping pixels in the image coordinate system to a geographic coordinate system to achieve georeferencing of the image.

[0021] Resampling and output: Finally, according to the geometric relationship between the rectified image and the original image, a new rectified image is generated by a resampling method (such as the nearest neighbor method, bilinear interpolation, or cubic convolution interpolation), and the output is saved.

[0022] Convert the rectified image coordinates into physical space coordinates.

[0023] Furthermore, S2 specifically includes the following steps:

[0024] Using the effective trajectory data of the traffic flow in S1, the time interval between vehicle arrivals and the headway between vehicles are simulated through a multi-level Monte Carlo model to predict the departure model of vehicles on the bridge.

[0025] A graph spatiotemporal neural network is used to predict vehicle information, abstracting the vehicle interaction relationships in the scene into a spatiotemporal undirected graph, where nodes represent vehicles and edges represent interactions between vehicles.

[0026] By constructing a multi-layer Monte Carlo model and graph spatiotemporal neural network, inputting vehicle information in the area and inferring the speed and model of newly arrived vehicles, a departure model that can reflect the actual traffic conditions is obtained.

[0027] Furthermore, the specific process of simulating the time interval between vehicle arrivals and the headway between vehicles using the multi-level Monte Carlo model includes:

[0028] A multi-level Monte Carlo model is used to simulate the time intervals and headway spacing of vehicles arriving at the bridge deck. Statistical methods and random sampling techniques are used to generate a series of possible vehicle arrival times and headway spacings based on historical traffic data. This allows the construction of a vehicle departure model that can reflect actual traffic conditions. The departure model can predict the probability distribution and time characteristics of vehicles arriving at the bridge deck under different traffic flows and time periods, providing accurate starting conditions for subsequent traffic flow simulations.

[0029] Furthermore, the specific process of abstracting the vehicle interaction relationship in the scene into a spatiotemporal undirected graph includes: after obtaining the distance between newly arrived vehicles, using the graph neural network to predict the vehicle information of the target vehicle, and abstracting the vehicle interaction relationship in the scene into a spatiotemporal undirected graph G(V,E), where V represents different vehicles and the graph edge E represents the interaction behavior between different vehicles, including the spatial edge set E s and the inter-frame edge set E f .

[0030] Furthermore, in the spatiotemporal undirected graph, the graph structure is represented by an adjacency matrix A={A0,A1}, both of which are of size V×V, where A0 is the self-connected unit matrix at the temporal level, and A1 is the adjacency matrix at the spatial level. Whether there is spatial adjacency is determined by whether the Euclidean distance between nodes is less than a set threshold.

[0031] Furthermore, S3 specifically includes the following steps:

[0032] Construction of trajectory prediction model: Using the effective traffic trajectory data formed in S1 and the vehicle embarkation model established in S2, a trajectory prediction model based on graph spatiotemporal neural network is constructed;

[0033] Implementation of random traffic flow simulation: The traffic flow data generated by the departure model is used as the initial input and put into the trajectory prediction model. The model predicts the data of vehicles in the driving area and merges the vehicle data with the data of newly arrived vehicles to form the input data at the next moment, realizing the micro-simulation of random traffic flow in the target area.

[0034] Furthermore, the trajectory prediction model unifies the vehicle's following and lane-changing behaviors into a trajectory prediction problem, using the same network structure as the departure module. The model's input features include the vehicle's lateral position x, longitudinal position y, speed, acceleration, and vehicle type, while the output is the vehicle's lateral position x' and longitudinal position y' at the next moment.

[0035] Furthermore, during the simulation, in the driving area, the vehicles that are still in the driving area in the prediction results are filtered out at each time step, and together with the vehicles entering the driving area from the departure area in the new frame, they are used as the input of the next prediction model. This process is repeated until the simulation time reaches the specified time length.

[0036] Compared with the prior art, the present invention has the following technical advantages:

[0037] 1) This paper combines a graph spatiotemporal neural network with a multi-level Monte Carlo model to provide a novel stochastic traffic flow simulation method that can more accurately predict vehicle driving behavior and traffic flow patterns. By simulating vehicle following and lane-changing behavior, this method generates traffic flow data that conforms to actual traffic conditions. This has important practical value for bridge design verification, traffic planning, and road safety maintenance.

[0038] 2) High portability and adaptability. Because the model is based on actual vehicle trajectory data, it can be easily applied to different bridge and road conditions, providing customized traffic flow simulation services for various transportation projects. Furthermore, this method can simulate extreme traffic conditions, providing more comprehensive safety assessments for engineering designs.

[0039] 3) Through refined driving simulation, this invention not only improves the accuracy of vehicle trajectory predictions but also enhances understanding of traffic flow dynamics. This significantly contributes to optimizing traffic flow, reducing congestion, and improving road efficiency. Furthermore, this method provides a powerful tool for traffic management agencies to better understand traffic flow characteristics and develop more effective traffic management strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1This is a schematic diagram of the overall process of the random traffic flow simulation method based on graph spatiotemporal neural network;

[0041] Figure 2 This is a schematic diagram of the surveillance video acquisition format;

[0042] Figure 3 This is a schematic diagram of perspective projection;

[0043] Figure 4 for multi-level Monte Carlo models and graph neural networks;

[0044] Figure 5 It is a vehicle driving flow chart;

[0045] Figure 6 Schematic diagram of the process of acquiring traffic data;

[0046] Figure 7 The following figure shows an example of simulation results. DETAILED DESCRIPTION

[0047] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Component models, material names, connection structures, control methods, algorithms, and other features not explicitly described in this technical solution are considered common technical features disclosed in the prior art.

[0048] Example 1

[0049] In general, constructing a vehicle load model for bridge fatigue reliability analysis is an important means of bridge design verification and safety maintenance. Problems that need to be solved in vehicle load model construction include traffic flow simulation, response analysis, model inference, etc., among which random traffic flow simulation is a key issue. However, traditional traffic micro-models cannot obtain traffic flows that are more consistent with actual speed change and lane change behaviors. The construction of a following model or lane change model based on a neural network ignores the connection between speed change and lane change behaviors. Therefore, this paper proposes a data-driven random traffic flow simulation method based on a graph spatiotemporal neural network, see Figure 1 The method of the present invention includes the following steps: 1) correcting, detecting, and tracking the bridge surface video captured by the camera to form effective trajectory data of the traffic flow; 2) establishing a vehicle departure model for getting on the bridge based on the multi-layer Monte Carlo model and the graph space-time neural network; 3) establishing a vehicle driving micro-simulation model based on the graph space-time neural network model. The present invention implements a traffic flow simulation method that conforms to actual speed change and lane change behaviors, and at the same time establishes a vehicle simulation model for vehicle speed change-lane change coupling. In addition, the proposed random traffic flow simulation method based on the graph space-time neural network has strong portability and important reference significance for the vehicle trajectory history analysis of bridges and other roads.

[0050] For the simulation of random vehicle flow in a specific range, in order to effectively consider the mutual following and lane changing relationship between vehicles and vehicle distribution, the first step is to obtain the historical data of the vehicle trajectory in the range. Specifically, several cameras are arranged in the area of interest to capture the vehicles on the bridge deck. The video acquisition form is shown in Figure 2 .

[0051] After obtaining the video, the image distortion is corrected by photogrammetry projection transformation. Specifically, the original image coordinates are converted into image coordinates that can correspond to the physical space relationship by using ground control points (GCPs) and total coordinate stations. The conversion form is shown in Figure 3 . This transformation process includes distortion correction and alignment of the image with reality, significantly improving accuracy and perspective. The corrected video is processed based on target detection and target tracking algorithms to record the current timestamp, vehicle number, geometric center position, and vehicle type, and on this basis, the vehicle speed and vehicle acceleration are calculated.

[0052] After obtaining the trajectory information of the vehicle queue, a multi-level Monte Carlo model and a graph neural network method are used for training, as shown in Figure 4 , to predict the vehicles that will arrive in the area. The traditional departure model uses Poisson distribution to obtain the headway between vehicles, and the vehicle speed and type are randomly assigned according to the probability density distribution of statistical data. Considering that the headway between two vehicles in a vehicle queue is affected by the congestion level of the traffic flow, the invention introduces a multi-level Monte Carlo model to obtain the headway data.

[0053] After obtaining the headway of the newly arrived vehicle, a graph neural network is used to predict the vehicle information of the vehicle. The interaction relationship between vehicles in the scene is abstracted into a space-time undirected graph G(V, E), where V represents different vehicles, and the graph edge E represents the interaction behavior between different vehicles, including the spatial edge set E s and the inter-frame edge set E f . The graph structure is represented by an adjacency matrix A = {A0, A1}, both of which are V x V. A0 is a self-connection unit matrix on the time level, and A1 is a spatial adjacency matrix. Whether two nodes are adjacent in space is determined by whether the Euclidean distance between them is less than a certain threshold. Based on this graph neural network, the vehicle information in the input area is used to infer the speed and type of the newly arrived vehicle.

[0054] The driving module of the vehicle unifies the following and lane changing behavior of the vehicle into trajectory prediction. The prediction model is also based on a graph space-time neural network, and the network structure is consistent with that of the departure module. However, when processing input data, the type and speed of the newly appeared vehicle do not need to be processed. The input features become lateral position x, longitudinal position y, speed, acceleration, and vehicle type, and the output becomes the lateral position x and longitudinal position y of the vehicle at the next time.

[0055] The traffic flow data generated by the departure module is input into the trajectory prediction model. The predicted data of vehicles in the driving area and the data of newly arrived vehicles together constitute the input data at the next moment. The final traffic flow simulation data is obtained by continuous looping. Figure 5 .

[0056] Based on the aforementioned multi-level Monte Carlo model, the departure graph neural network model, and the driving graph neural network model, a random traffic flow simulation is performed in the target area. After determining the initial state, this initial state is input into the departure model to generate several sets of departure data. This departure data is placed in a buffer within the departure area and then fed into the driving model at a time step to simulate random traffic flow within the area. Within the driving area, vehicles that remain in the driving area according to the prediction results are filtered out at each time step. These vehicles, along with the vehicles entering the driving area from the departure area in the next frame, form the input to the next prediction model. This cycle continues until the simulation reaches the specified time.

[0057] Application Example 1

[0058] To verify and analyze the proposed random traffic flow simulation method, video data of the bridge deck of a large-span suspension bridge was used as an example. Data processing was performed according to the above process. Through result analysis and efficiency calculation, the feasibility of the proposed method and its advantages of speed and accuracy were verified.

[0059] Data processing

[0060] Based on the video data of the bridge deck of a long-span suspension bridge, the surveillance camera is placed in the middle of the bridge tower beam and shoots 1920x1080 pixel video at a speed of 25 frames per second from a height of 85 meters. Image distortion is corrected by photogrammetric projection transformation. The coordinates are converted into a three-dimensional frame using ground control points (GCPs) and a total coordinate station. This homologous transformation process includes correcting distortion and aligning the image with reality, which significantly improves accuracy and perspective. The corrected video is then subjected to Yolov8-based target detection and Deepsort-based target tracking, recording the current timestamp, vehicle number, geometric center position, and vehicle type, and on this basis calculating the vehicle speed and vehicle acceleration. For a schematic diagram of the traffic data acquisition process, see Figure 6 .

[0061] Random traffic generation

[0062] This section uses 400 minutes of actual traffic data collected to train the spatiotemporal graph neural network for predicting vehicle model and speed, respectively. 100 minutes of data is used as the validation and test sets. Features include the vehicle's horizontal coordinate X and vertical coordinate Y, an adjacent distance of 20 meters, a maximum number of 40 nodes, an input step of 4 frames, a future step of 1 frame, and an interval of 0.5 seconds per frame. For a frame at time t in the actual traffic data, if there is a new vehicle in that frame, data from t-3 to t+1 is extracted and added to the training data. The vehicle model or speed of each vehicle when it first appears is set to 0 to predict the actual vehicle model or speed. The average inference time for sample speed is 3.6ms, and the average inference time for vehicle model is 3.4ms.

[0063] Then, headway distributions are obtained through various methods, converted into vehicle positions, and fed into a spatiotemporal graph neural network of vehicle type and speed, resulting in simulated initial vehicle data containing vehicle type, speed, and position. Headway is obtained using two methods: Monte Carlo sampling based on a fitted headway probability density function, and constructing a multi-layer Monte Carlo model. The multi-layer Monte Carlo model correlates the headway distances between vehicles at a given moment, which is more consistent with reality. The 2nth layer contains the lateral distance nodes between the nth and n+1th vehicles, and the 2n+1th layer contains the longitudinal distance nodes between the nth and n+1th vehicles. Each layer of the lateral distance layer has 10 nodes, with a spacing of 200 pixels between adjacent nodes. The kth node in the lateral distance layer indicates a lateral position between (k-1)*200 and k*200 pixels. The kth node in the longitudinal distance layer has 3 nodes per layer, with the kth node indicating the vehicle is in the kth lane.

[0064] Four consecutive random frames are selected as the initial input. Each departure lasts 800 frames. The two methods simulate the departure of about 30,000 vehicles each. The results are shown in Table 1.

[0065] Table 1 Statistics of vehicle speed and vehicle type under different departure methods

[0066]

[0067] As shown in Table 1, the proportion of each vehicle type, mean speed, and speed variance obtained by Multi-Layer Monte Carlo are closer to the actual situation.

[0068] Random traffic flow

[0069] The vehicle trajectory spatiotemporal graph neural network is trained using actual traffic data. The learning rate is set to 1×10 -4The batch size was 256. The neighbor distance was 20 meters, the maximum number of nodes was 100, the input step was 8 frames, the future step was 1 frame, and the interval between frames was 0.5 seconds. Various input features were tested, and the training results are shown in Table 2. When the features were X (horizontal axis), Y (vertical axis), vehicle type, speed, and acceleration, the loss, MAE, and RMSE were all minimized, so they were selected as the final features. The average inference time for this model sample was 4.5 milliseconds.

[0070] Table 1 Minimum training loss and accuracy of different features

[0071]

[0072] Random traffic flow simulation

[0073] Fifty sets of departure data from the departure model were input into the driving model. The departure area was 200 meters long and the driving area was 700 meters long. The simulation was repeated 50 times, each lasting 800 frames, with a 0.5-second interval between each frame. The X value of the departure section was set to 1800. Vehicles with a horizontal coordinate greater than this value in the departure data were considered to have entered the departure area. Vehicles in the departure area traveled at a constant speed and entered the driving area after a lateral displacement greater than 200 meters. Initially, frames 1 to 8 of the departure data were acquired and input into the trajectory prediction model to generate a prediction result. Vehicles still in the driving area from the prediction results were then filtered out and combined with the vehicles entering the driving area from the departure area in the next frame to form the input for the next prediction model. This cycle continued until the simulation reached the specified time.

[0074] Figure 7 This is an example of simulation results. The green blocks represent cars and the blue blocks represent trucks. Some vehicles show obvious lane-changing behavior, which shows that the system is capable of realizing micro-operating condition simulation.

[0075] The statistical characteristics of the headway distances of the measured data and the simulated data are shown in Table 2. The mean and standard deviation of the headway distances in each lane are relatively close.

[0076] Table 2 Random traffic flow simulation headway table

[0077]

[0078] The above description of the embodiments is intended to facilitate understanding and use of the invention by those skilled in the art. It will be apparent that those skilled in the art can readily make various modifications to these embodiments and apply the general principles described herein to other embodiments without requiring inventive effort. Therefore, the present invention is not limited to the above-described embodiments. Improvements and modifications made by those skilled in the art based on the disclosure of the present invention, without departing from the scope of the present invention, should be within the scope of protection of the present invention.

Claims

1. A random traffic flow simulation method based on graph spatiotemporal neural network, characterized in that: The following steps are involved: S1: Bridge surface videos captured by cameras are corrected, detected, and tracked to generate effective trajectory data for traffic. S2: Using the effective trajectory data of the traffic flow formed in S1, a vehicle departure model for entering the bridge is established based on the multi-layer Monte Carlo model and the graph spatiotemporal neural network; S3: Using the effective trajectory data of the traffic flow formed in S1 and the vehicle departure model established in S2, a vehicle driving micro-simulation model is established based on the graph spatiotemporal neural network model to realize the random traffic flow simulation on the bridge deck; S2 specifically includes the following steps: Using the effective trajectory data of the traffic flow in S1, the time interval between vehicle arrivals and the headway between vehicles are simulated through a multi-level Monte Carlo model to predict the departure model of vehicles on the bridge. A graph spatiotemporal neural network is used to predict vehicle information, abstracting the vehicle interaction relationships in the scene into a spatiotemporal undirected graph, where nodes represent vehicles and edges represent the interactions between vehicles. By using a multi-layer Monte Carlo model and graph spatiotemporal neural network, we input vehicle information in the area and infer the speed and type of newly arrived vehicles, thus obtaining a dispatch model that reflects the actual traffic conditions. In S3, the following steps are specifically included: Construction of trajectory prediction model: Using the effective traffic trajectory data formed in S1 and the vehicle embarkation model established in S2, a trajectory prediction model based on graph spatiotemporal neural network is constructed; Implementation of random traffic flow simulation: The traffic flow data generated by the departure model is used as the initial input and fed into the trajectory prediction model. The model predicts the data of vehicles in the driving area and merges this data with the data of newly arrived vehicles to form the input data for the next moment, thus realizing microscopic simulation of random traffic flow in the target area. The trajectory prediction model unifies the vehicle's following and lane-changing behaviors into a trajectory prediction problem, using the same network structure as the departure module. The model's input features include the vehicle's lateral position x, longitudinal position y, speed, acceleration, and vehicle type, and the output is the vehicle's lateral position x' and longitudinal position y' at the next moment. During the simulation, in the driving area, vehicles that are still in the driving area in the prediction results are filtered out at each time step and used as the input of the next prediction model together with the vehicles entering the driving area from the departure area in the new frame. This cycle is repeated until the simulation time reaches the specified time length.

2. The random traffic flow simulation method based on graph spatiotemporal neural network according to claim 1 is characterized in that: S1 specifically includes the following steps: Video rectification: Correct image distortion through photogrammetric projection transformation using ground control points and a global coordinate station; Vehicle detection and tracking: The corrected video is processed based on target detection and tracking algorithms, recording the current timestamp, vehicle number, geometric center position, and vehicle type, and calculating vehicle speed and acceleration. The acquired data together form effective trajectory data for the vehicle flow.

3. The random traffic flow simulation method based on graph spatiotemporal neural network according to claim 2 is characterized in that: The video correction process uses ground control points and a total coordinate station to perform photogrammetric projection transformation to correct image distortion and convert the original image coordinates into image coordinates of physical space relationships, specifically including: The original image is distorted using photogrammetry methods using ground control points and a total coordinate station; Convert the rectified image coordinates into physical space coordinates.

4. The random traffic flow simulation method based on graph spatiotemporal neural network according to claim 1 is characterized in that: The specific process of simulating the time interval and headway between vehicles using the multi-level Monte Carlo model includes: A multi-level Monte Carlo model is used to simulate the time intervals and headway spacing of vehicles arriving at the bridge deck. Statistical methods and random sampling techniques are used to generate a series of possible vehicle arrival times and headway spacings based on historical traffic data. This allows the construction of a vehicle departure model that can reflect actual traffic conditions. The departure model can predict the probability distribution and time characteristics of vehicles arriving at the bridge deck under different traffic flows and time periods, providing accurate starting conditions for subsequent traffic flow simulations.

5. The random traffic flow simulation method based on graph spatiotemporal neural network according to claim 1 is characterized in that: The specific process of abstracting the vehicle interaction relationship in the scene into a spatiotemporal undirected graph includes: after obtaining the distance between newly arrived vehicles, using the graph neural network to predict the vehicle information of the target vehicle, and abstracting the vehicle interaction relationship in the scene into a spatiotemporal undirected graph G(V, E), where V represents different vehicles and the graph edge E represents the interaction behavior between different vehicles, including the spatial edge set and inter-frame edge sets .

6. The random traffic flow simulation method based on graph spatiotemporal neural network according to claim 5 is characterized in that: In the spatiotemporal undirected graph, the graph structure adopts the adjacency matrix Indicates that both sizes are , is the self-connected identity matrix at the time level, It is the adjacency matrix at the spatial level. Whether the space is adjacent is determined by whether the Euclidean distance between the nodes is less than the set threshold.

Images