Urban traffic simulation model evaluation method and system based on data driving

Through the data-driven urban traffic simulation model, camera detection and DeepSORT algorithm are used to perform multimodal data correlation, solving the problems of long test cycles and high cost of autonomous driving, and achieving safe and efficient simulation verification.

CN120544101APending Publication Date: 2025-08-26NINGBO INST OF MATERIALS TECH & ENG CHINESE ACAD OF SCI
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202510657623.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-05-21
Publication Date
2025-08-26

AI Technical Summary

Technical Problem

The test cycle of autonomous driving technology is long, costly and has great safety risks, making it difficult to effectively verify in complex urban traffic scenarios.

Method used

The data-driven urban traffic simulation model is used to detect the agent target in real time through the camera, and multimodal data correlation and spatiotemporal consistency modeling is performed by combining the DeepSORT algorithm and Kalman filtering. The dictionary evaluation system is used to quantify the simulation results and generate a virtual test report.

Benefits of technology

Shorten the simulation test cycle, reduce costs, avoid the safety risks of road tests, and ensure the objective fidelity of simulation results through multi-dimensional quantitative indicators.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120544101A_ABST
    Figure CN120544101A_ABST
Patent Text Reader

Abstract

The invention discloses an urban traffic simulation model evaluation method and system based on data driving. The method comprises the following steps: detecting and tracking an intelligent body target in a video stream in real time through a camera, and outputting a target category, a position and a tracking identifier; wherein the intelligent body targets are vehicles and pedestrians; performing multi-modal data association and time-space consistency modeling on the motion trail of the intelligent body target; through a data-driven optimization algorithm, a road scene simulation technology is fused, and complex traffic test conditions are dynamically restored; carrying out authenticity measurement on a simulation result by utilizing a dictionary evaluation system, and quantifying speed distribution, density distribution and track error indexes; the simulation data are fed back to the test platform, the model parameters are optimized, and a virtual test report is generated. According to the invention, road testing is partially replaced and complemented with the road testing, so that the testing period can be greatly shortened, the testing cost can be reduced, and the life safety of testing personnel can be protected.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This technology relates to the fields of intelligent transportation and autonomous driving technology, and specifically to a data-driven urban traffic simulation model evaluation method and system. Background Art

[0002] In recent years, autonomous driving technology has made significant progress. However, its development and application face numerous challenges. While real-world road testing can capture real-world traffic data and verify the technology's effectiveness, it also presents significant challenges, such as long testing cycles, high costs, and significant safety risks.

[0003] On the other hand, complex and ever-changing urban traffic scenarios involve a variety of traffic participants, including motor vehicles, non-motor vehicles, and pedestrians, as well as varying road types, traffic signals, weather conditions, and other environmental factors. This places extremely high demands on the reliability and robustness of autonomous driving systems. To ensure the safety and effectiveness of autonomous driving technology under these complex conditions, extensive testing and verification is required. Therefore, an urban traffic simulation model is needed to simulate realistic traffic conditions and provide strong support for the research, development, optimization, and verification of autonomous driving technology. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention addresses the problems of long test cycles, high test costs, and poor tester safety during autonomous driving tests. It proposes a data-driven urban traffic simulation model evaluation method and system. This model partially replaces road tests and complements road tests, thereby solving the technical problems raised in the background technology.

[0005] In order to achieve the aforementioned object of the invention, the present invention adopts the following scheme:

[0006] One aspect of the present invention provides a data-driven urban traffic simulation model evaluation method, comprising:

[0007] Step S1: Detect and track intelligent targets in the video stream in real time through a camera, and output the target category, location, and tracking identifier; wherein the intelligent targets are vehicles and pedestrians;

[0008] Step S2: Perform multimodal data association and spatiotemporal consistency modeling on the target motion trajectory of the intelligent agent;

[0009] Step S3: Dynamically restore complex traffic test conditions by integrating data-driven optimization algorithms with road scene simulation technology;

[0010] Step S4: Use the dictionary evaluation system to measure the authenticity of the simulation results and quantify the velocity distribution, density distribution and trajectory error indicators;

[0011] Step S5: Feedback the simulation data to the test platform, optimize the model parameters and generate a virtual test report.

[0012] In one embodiment, the data driving in step S3 includes obtaining the real-world speed and position of the intelligent target and mapping it on a high-definition street view road, thereby simulating complex traffic test conditions.

[0013] In one embodiment, step S1: detecting and tracking an intelligent target in a video stream in real time using a camera, and outputting the target category, location, and tracking identifier, specifically includes:

[0014] Identify road scenes and extract key visual information. Pre-train the YOLOv8 network using a labeled dataset. Apply the pre-trained YOLOv8 network to extract key road feature information to identify intelligent targets, output the center coordinates (x, y) of their bounding boxes and category labels, and provide a spatial position reference for subsequent motion analysis.

[0015] In one embodiment, step S2 performs multimodal data association and spatiotemporal consistency modeling on the detected target motion trajectory, specifically including:

[0016] The DeepSORT algorithm is adopted to extract and fuse the target's appearance features and motion state through the deep learning-based ReID network, build a multimodal data association model, and establish an ID binding mechanism to ensure the independent calculation of the speed of different intelligent targets. CNN is used to extract the target's appearance features and generate a high-dimensional embedding vector. The identity consistency of the target across frames is measured by cosine similarity. The Hungarian algorithm is used to optimize the matching weights of detection and trajectory to ensure smooth transition of the intelligent target's trajectory between consecutive frames and reduce jitter or mutation, thereby achieving multimodal data association and spatiotemporal consistency modeling.

[0017] In one embodiment, the center coordinates (x t ,y t ) and (x t+Δt ,y t+Δt ), calculate the displacement

[0018] Determine the time interval Δt=1 / FPS based on the video frame rate FPS;

[0019] Through camera calibration and perspective transformation, the pixel displacement is mapped to the actual spatial scale, and the speed display is converted to km / h;

[0020] The Savitzky-Golay filter smoothing algorithm is combined to suppress noise and estimate the instantaneous velocity: V = Δd / Δt.

[0021] In one embodiment, in step S4, a dictionary-based evaluation system is used to measure the authenticity of the simulation results, including:

[0022] Absolute mean error of speed:

[0023]

[0024] Where SE is the absolute mean error of the agent's target speed, SE ≥ O, the smaller the value, the closer the simulation is to the reference data, T is the total number of simulation time steps, t is the current step size, t∈{0,1,2,...,T}, L and Q are the sets of rows L and columns Q of the network partition, and the traffic area is divided into 1×q unit grids 1∈{1,2,...,Lmax},q∈{1,2,...,Qmax}, is the average speed of the simulated agent target i in the l×q cells, is the average speed of agent target i in l×q cells;

[0025] Absolute mean error of density:

[0026]

[0027] The road is divided into l× q cells, each cell is 1 meter long and wide. In each cell, we calculate the density of the agent's target to calculate the density distribution of the agent's target; is the number of simulated agent targets in the l×q cells, is the density of the real agent target in l×q cells;

[0028] Track length error:

[0029]

[0030] Among them, FDE is the trajectory length indicator between the simulated group and the real group, N is the total number of agent goals, T is the total simulation step length, is the position of the real agent target at time t+1, is the position of the real agent target at time t. is the position of the simulation agent target at time t+1.

[0031] Trajectory cumulative absolute difference:

[0032]

[0033] Where N is the total number of agent goals, T is the total simulation step length, is the position of the real agent target at time t, is the position of the simulated agent target at time t.

[0034] In one embodiment, in step S5, SUMO and CARLA are used as the test platform for the intelligent agent target simulation data. The pixel coordinates are mapped to the physical coordinate system through the homography matrix to generate real-world coordinates (X, Y). These coordinates are mapped to the established SUMO network, and a virtual test report is output to ensure that all indicators remain within the specified range.

[0035] Another aspect of the present invention provides a data-driven urban traffic simulation model evaluation system, comprising:

[0036] The intelligent target detection and real-time tracking module is used to detect and track intelligent targets in the video stream in real time through the camera, and output the target category, location and tracking identifier; the intelligent targets are vehicles and pedestrians;

[0037] Multimodal data association and trajectory modeling module, used to perform multimodal data association and spatiotemporal consistency modeling on the target motion trajectory of the intelligent agent;

[0038] Dynamic scenario generation and optimization module, which uses data-driven optimization algorithms and integrates road scenario simulation technology to dynamically restore complex traffic test conditions;

[0039] The authenticity assessment and index quantification module is used to measure the authenticity of simulation results using a dictionary-based evaluation system and quantify the velocity distribution, density distribution and trajectory error indicators.

[0040] The virtual test report generation module is used to feed back simulation data to the test platform, optimize model parameters and generate virtual test reports.

[0041] In one embodiment, data driving includes obtaining the real-world speed and position of the intelligent target and mapping it onto a high-definition street view road, thereby simulating complex traffic test conditions.

[0042] In one embodiment, the intelligent object detection and real-time tracking module further includes:

[0043] Visual feature extraction module, used to identify road scenes and extract key visual information;

[0044] The target recognition and tracking module is used to pre-train the YOLOv8 network using a labeled dataset. The pre-trained YOLOv8 network is then used to extract key road feature information to identify intelligent targets. The module then outputs the center coordinates (x, y) of the bounding box and the category label, providing a spatial position reference for subsequent motion analysis.

[0045] In one embodiment, the multimodal data association and trajectory modeling module further includes:

[0046] The multimodal data association model construction module is used to use the DeepSORT algorithm to extract and fuse the target appearance features and motion state through the deep learning-based ReID network, build a multimodal data association model, and establish an ID binding mechanism to ensure the independent calculation of the speed of different intelligent targets;

[0047] The trajectory consistency module is used to use CNN to extract the apparent features of the target, generate a high-dimensional embedding vector, measure the identity consistency of the target across frames through cosine similarity, and optimize the matching weight of detection and trajectory through the Hungarian algorithm to ensure smooth transition of the trajectory of the intelligent target between consecutive frames and reduce jitter or mutation, thereby achieving multimodal data association and spatiotemporal consistency modeling.

[0048] In one embodiment, the center coordinates (x t ,y t ) and (X t+Δt ,y t+Δt ), calculate the displacement

[0049] The time interval Δt=1 / FPS is determined based on the video frame rate FPS.

[0050] Through camera calibration and perspective transformation, the pixel displacement is mapped to the actual spatial scale, and the speed display is converted into km / h.

[0051] The Savitzky-Golay filter smoothing algorithm is combined to suppress noise and estimate the instantaneous velocity: V = Δd / Δt.

[0052] In one embodiment, the authenticity metric includes:

[0053] Absolute mean error of speed:

[0054]

[0055] Where SE is the absolute mean error of the agent's target speed, SE ≥ 0, the smaller the value, the closer the simulation is to the reference data, T is the total number of simulation time steps, t is the current step size, t∈{0,1,2,...,T}, L and Q are the sets of rows L and columns Q of the network partition, and the traffic area is divided into 1×q unit grids 1∈{1,2,...,Lmax},q∈{1,2,...,Qmax}, is the average speed of the simulated agent target i in the l×q cells, is the average speed of agent target i in l×q cells;

[0056] Absolute mean error of density:

[0057]

[0058] The road is divided into l×q cells, each of which is 1 meter long and 1 meter wide. In each cell, we calculate the density of the agent's target to calculate the density distribution of the agent's target. is the number of simulated agent targets in the l×q cells, is the density of the real agent target in l×q cells;

[0059] Track length error:

[0060]

[0061] Among them, FDE is the trajectory length indicator between the simulated group and the real group, N is the total number of agent goals, T is the total simulation step length, is the position of the real agent target at time t+1, is the position of the real agent target at time t. is the position of the simulation agent target at time t+1.

[0062] Trajectory cumulative absolute difference:

[0063]

[0064] Where N is the total number of agent goals, T is the total simulation step length, is the position of the real agent target at time t, is the position of the simulated agent target at time t.

[0065] In one embodiment, the virtual test report generation module further includes: using SUMO and CARLA as the test platform for intelligent agent target simulation data, mapping pixel coordinates to the physical coordinate system through the homography matrix, generating real-world coordinates (X, Y), and mapping them in the established SUMO network, outputting a virtual test report, and ensuring that various indicators remain within the specified range.

[0066] Compared with the prior art, the present invention has at least the following advantages:

[0067] (1) The present invention integrates the DeepSORT algorithm through YOLOv8, combines appearance feature matching with Kalman filter prediction, and maintains high accuracy of identity in occlusion scenarios.

[0068] (2) The present invention constructs a multimodal joint optimization model by integrating visual detection (YOLOv8), motion prediction (Kalman filter) and scene semantics (lane topology, traffic light status).

[0069] (3) The present invention establishes a dictionary-based evaluation system and predefines multi-dimensional quantitative indicators (absolute mean error of velocity distribution <0.02, cumulative absolute difference index of trajectory <0.9m) to support objective fidelity scoring.

[0070] (4) The present invention realizes a virtual test closed loop through the SUMO platform, is compatible with complex urban road networks (intersections, ramps, non-motorized vehicle lanes), and supports in-loop testing of autonomous driving algorithms. Compared with real vehicle testing, it shortens the simulation cycle, reduces costs, and avoids the safety risks of road testing. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0072] Figure 1 This is a module composition diagram of a data-driven urban traffic simulation system provided by a typical embodiment of the present invention. DETAILED DESCRIPTION

[0073] To make the objectives, technical solutions, and advantages of the present invention more apparent, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings. Examples of these preferred embodiments are illustrated in the accompanying drawings. The embodiments of the present invention shown in and described with reference to the accompanying drawings are merely exemplary, and the present invention is not limited to these embodiments.

[0074] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0075] like Figure 1 As shown, the present invention provides a data-driven urban traffic simulation model evaluation method, comprising:

[0076] Step S1: Detect and track intelligent targets in the video stream in real time through a camera, and output the target category, location, and tracking identifier; wherein the intelligent targets are vehicles and pedestrians;

[0077] Step S2: Perform multimodal data association and spatiotemporal consistency modeling on the target motion trajectory of the intelligent agent;

[0078] Step S3: Dynamically restore complex traffic test conditions by integrating data-driven optimization algorithms with road scene simulation technology;

[0079] Step S4: Use the dictionary evaluation system to measure the authenticity of the simulation results and quantify the velocity distribution, density distribution and trajectory error indicators;

[0080] Step S5: Feedback the simulation data to the test platform, optimize the model parameters and generate a virtual test report.

[0081] In this example, the data drive in step S3 includes the real-world speed and position of the intelligent target, and maps it into a high-definition street view road, thereby simulating complex traffic conditions.

[0082] In this example, the road scene in step S1 is identified and key visual information is extracted. The YOLOv8 network is pre-trained using a labeled dataset. The pre-trained YOLOv8 network is used to extract key road feature information to identify the intelligent target, and the center coordinates (x, y) and category label of its bounding box are output to provide a spatial position reference for subsequent motion analysis.

[0083] In this example, the DeepSORT algorithm is used in step S2. By fusing the target appearance features (extracted by the ReID network based on deep learning) and the motion state (displacement, speed and other parameters predicted by the Kalman filter), a multimodal data association model is constructed, and an ID binding mechanism is established to ensure that the speeds of different intelligent targets are calculated independently. CNN is used to extract the target's appearance features and generate a high-dimensional embedding vector. The identity consistency of the target across frames is measured by cosine similarity to solve the problem of intelligent target identity switching caused by occlusion or deformation, so that the speed continuity is maintained when the target disappears briefly. The Hungarian Algorithm is also used to optimize the matching weights of detection and trajectory to ensure a smooth transition of the target's trajectory between consecutive frames and reduce jitter or mutation. This achieves multimodal data association and spatiotemporal consistency modeling.

[0084] Furthermore, in order to solve the problem of disappearance of the intelligent target identity ID caused by occlusion and deformation, a ReID network is introduced. By using the ReID network to extract the high-dimensional embedding vectors of the detection target and the trajectory target, the cosine similarity is calculated:

[0085] C appearance =1-cosine similarity(f dectect -f track )

[0086] Among them, f dectect and f track are the feature vectors of the detected target and the track target respectively.

[0087] Furthermore, to solve the motion matching problem, a Kalman filter was introduced to maintain velocity continuity even when the target disappears or deforms briefly. The Kalman filter predicts the position and velocity of the current frame based on the state of the previous frame, as shown in the following formula:

[0088] x t =F·x t-1 +w t

[0089] Where F is the state transition matrix and wt is the process noise. The prediction value is corrected using the detection result of the current frame, and the uncertainty is adjusted through the covariance matrix.

[0090] Furthermore, in motion metrics, the Mahalanobis distance is used to measure the matching degree between the detection box and the prediction box, while taking into account the covariance matrix (and uncertainty). The formula of the Mahalanobis distance is as follows:

[0091]

[0092] in, is the innovation covariance matrix.

[0093] To reduce jitter during target detection, the Hungarian algorithm is used. It is a polynomial-time algorithm for solving the minimum weight matching problem on bipartite graphs. In multi-target tracking, its core goal is to optimally correlate the detection results of the current frame with existing trajectories, ensuring cross-frame target identity continuity and trajectory smoothness. By fusing the target's appearance features and motion state, a comprehensive cost matrix is ​​constructed:

[0094] C ij =λ·C appearance +(1-λ)·C motion

[0095] When the target is not occluded, increase the appearance weight λ. When the target is occluded or the lighting changes drastically, reduce the appearance weight λ and rely on motion prediction.

[0096] In this example, step S3 first requires building a high-definition road simulation model in SUMO. This involves defining lanes, intersections, traffic lights, and speed limits based on actual road data or manual drawing, and generating the road using a network file (.net.xml) to ensure lane connectivity and traffic light phase conflicts. The simulation duration is also defined and dynamically controlled through the TraCI interface.

[0097] Furthermore, according to the center coordinates (x t ,y t ) and (x t+Δt ,y t+Δt ), calculate the displacement (pixel units);

[0098] Determine the time interval Δt=1 / FPS (seconds) based on the video frame rate (FPS);

[0099] Through camera calibration and perspective transformation, the pixel displacement is mapped to the actual spatial scale, and the speed display is converted to km / h;

[0100] The Savitzky-Golay filter smoothing algorithm is combined to suppress noise and estimate the instantaneous velocity: V = Δd / Δt.

[0101] Furthermore, in terms of camera calibration, the Zhang Zhengyou calibration method is used to obtain the camera internal parameters (focal length f x , f y 、Main point c x ,,c y ) and distortion coefficients to correct image distortion.

[0102] Select at least 4 sets of landmark points (such as lane intersections) on the ground plane of the scene and measure their physical coordinates. Solve the homography matrix H through SVD decomposition to satisfy:

[0103]

[0104] Map the target center pixel coordinates to the physical coordinate system and output the X, Y position (unit: meter). Synchronize with the simulation clock based on the video timestamp to ensure timing consistency.

[0105] Furthermore, when combining the Savitzky-Golay filter smoothing algorithm to suppress noise, the collected trajectory data is first preprocessed to ensure a continuous time series with no missing values. Linear interpolation is then used to fill in anomalous or missing data points. Subsequently, filtering parameters—window size and polynomial order—are selected based on the characteristics of the traffic scenario. In densely packed traffic, a larger window size is used to enhance noise suppression; in free-flow scenarios, a smaller window size is chosen to preserve details of lane changes or accelerations. The polynomial order is typically set to 2 or 3 to balance smoothing effectiveness with signal fidelity. During the filtering process, a local polynomial fit is performed on each data point, taking its neighboring points as the center. The smoothed trajectory coordinates or velocity values ​​are then calculated. A mirror padding strategy is applied to the data boundaries to prevent distortion of the leading and trailing data due to window crossings. To address dynamic traffic conditions, parameters are dynamically adjusted based on real-time noise assessment, temporarily increasing the window size in areas of high jitter. Finally, a dictionary-based evaluation system was used to measure the authenticity of the simulation results and conduct trajectory visualization comparison to verify the filtering effect, ensuring that the overall position deviation was controlled within 0.1 meters, thereby providing high-fidelity, low-noise input data for the simulation model.

[0106] Furthermore, the pre-processed agent target data is stored in the form of a dictionary and input into the constructed SUMO network, and the SUMO planar road network is mapped to the CARLA three-dimensional world coordinate system to complete the joint simulation of SUMO and CARLA.

[0107] In this example, the fourth step uses a dictionary-based evaluation system to measure the authenticity of the simulation results, including:

[0108] Absolute mean error of speed:

[0109]

[0110] Where SE is the absolute mean error of the agent's target speed, SE ≥ 0, the smaller the value, the closer the simulation is to the reference data, T is the total number of simulation time steps, t is the current step size, t∈{0, 1, 2, ..., T}, L and Q are the sets of rows L and columns Q of the network partition, and the traffic area is divided into 1×q unit grids 1∈{1, 2, ..., Lmax}, q∈{1, 2, ..., Qmax}, is the average speed of the simulated agent target i in the l×q cells, is the average speed of agent target i in l×q cells;

[0111] Absolute mean error of density:

[0112]

[0113] The road is divided into l×q cells, each cell is 1 meter long and wide. In each cell, we calculate the density of the agent's target to calculate the density distribution of the agent's target. is the number of simulated agent targets in the l×q cells, is the density of the real agent target in l×q cells;

[0114] Track length error:

[0115]

[0116] Among them, FDE is the trajectory length indicator between the simulated group and the real group, N is the total number of agent goals, T is the total simulation step length, is the position of the real agent target at time t+1, is the position of the real agent target at time t. is the position of the simulation agent target at time t+1.

[0117] Trajectory cumulative absolute difference:

[0118]

[0119] Where N is the total number of agent goals, T is the total simulation step length, is the position of the real agent target at time t, is the position of the simulated agent target at time t.

[0120] The system uses an automated interface to perform multi-dimensional comparative analysis between the simulation outputs of SUMO and CARLA and real-world scene data. In crowd simulation results, the cumulative absolute difference index of trajectories is less than 0.9m, the trajectory length index is less than 0.5m, the absolute average error of velocity distribution is less than 0.02, and the absolute average error of density distribution is less than 0.01m. In vehicle simulation results, the absolute average error of velocity distribution is less than 0.01, and the absolute average error of density distribution is less than 0.01m.

[0121] like Figure 1 As shown, the present invention provides a data-driven urban traffic simulation model evaluation system, comprising:

[0122] The intelligent target detection and real-time tracking module is used to detect and track intelligent targets in the video stream in real time through the camera, and output the target category, location and tracking identifier; the intelligent targets are vehicles and pedestrians;

[0123] Multimodal data association and trajectory modeling module, used to perform multimodal data association and spatiotemporal consistency modeling on the target motion trajectory of the intelligent agent;

[0124] Dynamic scenario generation and optimization module, which uses data-driven optimization algorithms and integrates road scenario simulation technology to dynamically restore complex traffic test conditions;

[0125] The authenticity assessment and index quantification module is used to measure the authenticity of simulation results using a dictionary-based evaluation system and quantify the velocity distribution, density distribution and trajectory error indicators.

[0126] The virtual test report generation module is used to feed back simulation data to the test platform, optimize model parameters and generate virtual test reports.

[0127] In one embodiment, data driving includes obtaining the real-world speed and position of the intelligent target and mapping it onto a high-definition street view road, thereby simulating complex traffic test conditions.

[0128] In one embodiment, the intelligent object detection and real-time tracking module further includes:

[0129] Visual feature extraction module, used to identify road scenes and extract key visual information;

[0130] The target recognition and tracking module is used to pre-train the YOLOv8 network using a labeled dataset. The pre-trained YOLOv8 network is then used to extract key road feature information to identify intelligent targets. The module then outputs the center coordinates (x, y) of the bounding box and the category label, providing a spatial position reference for subsequent motion analysis.

[0131] In one embodiment, the multimodal data association and trajectory modeling module further includes:

[0132] The multimodal data association model construction module is used to use the DeepSORT algorithm to extract and fuse the target appearance features and motion state through the deep learning-based RelD network, build a multimodal data association model, and establish an ID binding mechanism to ensure the independent calculation of the speed of different intelligent targets;

[0133] The trajectory consistency module is used to use CNN to extract the apparent features of the target, generate a high-dimensional embedding vector, measure the identity consistency of the target across frames through cosine similarity, and optimize the matching weight of detection and trajectory through the Hungarian algorithm to ensure smooth transition of the trajectory of the intelligent target between consecutive frames and reduce jitter or mutation, thereby achieving multimodal data association and spatiotemporal consistency modeling.

[0134] Furthermore, in order to solve the problem of disappearance of the intelligent target identity ID caused by occlusion and deformation, a ReID network is introduced. By using the ReID network to extract the high-dimensional embedding vectors of the detection target and the trajectory target, the cosine similarity is calculated:

[0135] C appearance =1-cosine similarity(f dectect -f track )

[0136] Among them, f dectect and f track are the feature vectors of the detected target and the track target respectively.

[0137] Furthermore, to solve the motion matching problem, a Kalman filter was introduced to maintain velocity continuity even when the target disappears or deforms briefly. The Kalman filter predicts the position and velocity of the current frame based on the state of the previous frame, as shown in the following formula:

[0138] x t =F·x t-1 +w t

[0139] Where F is the state transition matrix and wt is the process noise. The prediction value is corrected using the detection result of the current frame, and the uncertainty is adjusted through the covariance matrix.

[0140] Furthermore, in motion metrics, the Mahalanobis distance is used to measure the matching degree between the detection box and the prediction box, while taking into account the covariance matrix (and uncertainty). The formula of the Mahalanobis distance is as follows:

[0141]

[0142] in, is the innovation covariance matrix.

[0143] To reduce jitter during target detection, the Hungarian algorithm is used. It is a polynomial-time algorithm for solving the minimum weight matching problem on bipartite graphs. In multi-target tracking, its core goal is to optimally correlate the detection results of the current frame with existing trajectories, ensuring cross-frame target identity continuity and trajectory smoothness. By fusing the target's appearance features and motion state, a comprehensive cost matrix is ​​constructed:

[0144] C ij =λ·C appearance +(1-λ)·C motion

[0145] When the target is not occluded, increase the appearance weight λ. When the target is occluded or the lighting changes drastically, reduce the appearance weight λ and rely on motion prediction.

[0146] In one embodiment, the center coordinates (x t ,y t ) and (x t+Δt ,y t+Δt ), calculate the displacement

[0147] The time interval Δt=1 / FPS is determined based on the video frame rate FPS.

[0148] Through camera calibration and perspective transformation, the pixel displacement is mapped to the actual spatial scale, and the speed display is converted into km / h.

[0149] The Savitzky-Golay filter smoothing algorithm is combined to suppress noise and estimate the instantaneous velocity: V = Δd / Δt.

[0150] Furthermore, in terms of camera calibration, the Zhang Zhengyou calibration method is used to obtain the camera internal parameters (focal length f x , f y 、Main point c x ,,c y ) and distortion coefficients to correct image distortion;

[0151] Select at least four sets of landmark points (such as lane intersections) on the ground plane of the scene, measure their physical coordinates, and solve the homography matrix H through SVD decomposition to satisfy:

[0152]

[0153] Map the target center pixel coordinates to the physical coordinate system, output the X, Y position (unit: meter), and synchronize with the simulation clock based on the video timestamp to ensure timing consistency.

[0154] Furthermore, when combining the Savitzky-Golay filter smoothing algorithm to suppress noise, the collected trajectory data is first preprocessed to ensure a continuous time series with no missing values. Linear interpolation is then used to fill in anomalous or missing data points. Subsequently, filtering parameters—window size and polynomial order—are selected based on the characteristics of the traffic scenario. In densely packed traffic, a larger window size is used to enhance noise suppression; in free-flow scenarios, a smaller window size is chosen to preserve details of lane changes or accelerations. The polynomial order is typically set to 2 or 3 to balance smoothing effectiveness with signal fidelity. During the filtering process, a local polynomial fit is performed on each data point, taking its neighboring points as the center. The smoothed trajectory coordinates or velocity values ​​are then calculated. A mirror padding strategy is applied to the data boundaries to prevent distortion of the leading and trailing data due to window crossings. To address dynamic traffic conditions, parameters are dynamically adjusted based on real-time noise assessment, temporarily increasing the window size in areas of high jitter. Finally, a dictionary-based evaluation system was used to measure the authenticity of the simulation results and conduct trajectory visualization comparison to verify the filtering effect, ensuring that the overall position deviation was controlled within 0.1 meters, thereby providing high-fidelity, low-noise input data for the simulation model.

[0155] Furthermore, the pre-processed agent target data is stored in the form of a dictionary and input into the constructed SUMO network, and the SUMO planar road network is mapped to the CARLA three-dimensional world coordinate system to complete the joint simulation of SUMO and CARLA.

[0156] In one embodiment, the authenticity metric includes:

[0157] Absolute mean error of speed:

[0158]

[0159] Where SE is the absolute mean error of the agent's target speed, SE ≥ 0, the smaller the value, the closer the simulation is to the reference data, T is the total number of simulation time steps, t is the current step size, t∈{0, 1, 2, ..., T}, L and Q are the sets of rows L and columns Q of the network partition, and the traffic area is divided into l×q unit grids l∈{1, 2, ..., Lmax}, q∈{1, 2, ..., Qmax}, is the average speed of the simulated agent target i in the l×q cells, is the average speed of agent target i in l×q cells;

[0160] Absolute mean error of density:

[0161]

[0162] The road is divided into l×q cells, each of which is 1 meter long and 1 meter wide. In each cell, we calculate the density of the agent's target to calculate the density distribution of the agent's target. is the number of simulated agent targets in the l×q cells, is the density of the real agent target in l×q cells;

[0163] Track length error:

[0164]

[0165] Among them, FDE is the trajectory length indicator between the simulated group and the real group, N is the total number of agent goals, T is the total simulation step length, is the position of the real agent target at time t+1, is the position of the real agent target at time t. is the position of the simulation agent target at time t+1.

[0166] Trajectory cumulative absolute difference:

[0167]

[0168] Where N is the total number of agent goals, T is the total simulation step length, is the position of the real agent target at time t, is the position of the simulated agent target at time t.

[0169] In one embodiment, the virtual test report generation module further includes: using SUMO and CARLA as the test platform for intelligent agent target simulation data, mapping pixel coordinates to the physical coordinate system through the homography matrix, generating real-world coordinates (X, Y), and mapping them in the established SUMO network, outputting a virtual test report, and ensuring that various indicators remain within the specified range.

[0170] The system uses an automated interface to perform multi-dimensional comparative analysis between the simulation outputs of SUMO and CARLA and real-world scene data. In crowd simulation results, the cumulative absolute difference index for trajectories is less than 0.9m, the trajectory length index is less than 0.5m, the absolute average error in velocity distribution is less than 0.02, and the absolute average error in density distribution is less than 0.01m. In vehicle simulation results, the absolute average error in velocity distribution is less than 0.01, and the absolute average error in density distribution is less than 0.01m.

[0171] It should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions of each embodiment can also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A data-driven urban traffic simulation model evaluation method, characterized in that: include: Step S1: Detect and track the intelligent target in the video stream in real time through the camera, and output the target category, location and tracking mark; The intelligent targets are vehicles and pedestrians; Step S2: Perform multimodal data association and spatiotemporal consistency modeling on the target motion trajectory of the intelligent agent; Step S3: Dynamically restore complex traffic test conditions by integrating data-driven optimization algorithms with road scene simulation technology; Step S4: Use the dictionary evaluation system to measure the authenticity of the simulation results and quantify the velocity distribution, density distribution and trajectory error indicators; Step S5: Feedback the simulation data to the test platform, optimize the model parameters and generate a virtual test report.

2. The data-driven urban traffic simulation model evaluation method according to claim 1 is characterized in that: The data driving in step S3 includes obtaining the real-world speed and position of the intelligent target and mapping it on a high-definition street view road, thereby simulating complex traffic test conditions.

3. The data-driven urban traffic simulation model evaluation method according to claim 2 is characterized in that: Step S1: detecting and tracking the intelligent target in the video stream in real time through the camera, and outputting the target category, location and tracking mark, specifically includes: Identify road scenes and extract key visual information. Pre-train the YOLOv8 network using a labeled dataset. Apply the pre-trained YOLOv8 network to extract key road feature information to identify intelligent targets, output the center coordinates (x, y) of their bounding boxes and category labels, and provide a spatial position reference for subsequent motion analysis.

4. The data-driven urban traffic simulation model evaluation method according to claim 1 is characterized in that: The step S2 performs multimodal data association and spatiotemporal consistency modeling on the motion trajectory of the detected target, specifically including: The DeepSORT algorithm is adopted to extract and fuse the target's appearance features and motion state through the deep learning-based ReID network, build a multimodal data association model, and establish an ID binding mechanism to ensure the independent calculation of the speed of different intelligent targets. CNN is used to extract the target's appearance features and generate a high-dimensional embedding vector. The identity consistency of the target across frames is measured by cosine similarity. The Hungarian algorithm is used to optimize the matching weights of detection and trajectory to ensure smooth transition of the intelligent target's trajectory between consecutive frames and reduce jitter or mutation, thereby achieving multimodal data association and spatiotemporal consistency modeling.

5. The data-driven urban traffic simulation model evaluation method according to claim 3 is characterized in that: According to the center coordinates (x t ,y t ) and (x t+Δt ,y t+Δt ), calculate the displacement Determine the time interval Δt=1 / FPS based on the video frame rate FPS; Through camera calibration and perspective transformation, the pixel displacement is mapped to the actual spatial scale, and the speed display is converted to km / h; The Savitzky-Golay filter smoothing algorithm is combined to suppress noise and estimate the instantaneous velocity: V = Δd / Δt.

6. The data-driven urban traffic simulation model evaluation method according to claim 1 is characterized in that: In step S4, a dictionary-based evaluation system is used to measure the authenticity of the simulation results, including: Absolute mean error of speed: Where SE is the absolute mean error of the agent's target speed, SE ≥ 0, the smaller the value, the closer the simulation is to the reference data, T is the total number of simulation time steps, t is the current step size, t∈{O, 1, 2, ..., T}, L and Q are the sets of rows L and columns Q of the network partition, and the traffic area is divided into 1×q unit grids 1∈{1, 2, ..., Lmax}, q∈{1, 2, ..., Qmax}, is the average speed of the simulated agent target i in the l×q cells, is the average speed of agent target i in l×q cells; Absolute mean error of density: The road is divided into l×q cells, each of which is 1 meter long and 1 meter wide. The density of the agent target is calculated in each cell to calculate the density distribution of the agent target. is the number of simulated agent targets in the l×q cells, is the density of the real agent target in l×q cells; Track length error: Among them, FDE is the trajectory length indicator between the simulated group and the real group, N is the total number of agent goals, T is the total simulation step length, is the position of the real agent target at time t+1, is the position of the real agent target at time t. is the position of the simulation agent target at time t+1. Cumulative absolute difference of trajectories: Where N is the total number of agent goals, T is the total simulation step length, is the position of the real agent target at time t, is the position of the simulated agent target at time t.

7. The data-driven urban traffic simulation model evaluation method according to claim 1 is characterized in that: In step S5, SUMO and CARLA are used as the test platform for the intelligent agent target simulation data. The pixel coordinates are mapped to the physical coordinate system through the homography matrix to generate real-world coordinates (x, Y). These coordinates are mapped to the established SUMO network and a virtual test report is output to ensure that all indicators remain within the specified range.

8. A data-driven urban traffic simulation model evaluation system, characterized in that: include: The intelligent target detection and real-time tracking module is used to detect and track intelligent targets in the video stream in real time through the camera, and output the target category, location and tracking mark; The intelligent targets are vehicles and pedestrians; Multimodal data association and trajectory modeling module, used to perform multimodal data association and spatiotemporal consistency modeling on the target motion trajectory of the intelligent agent; Dynamic scenario generation and optimization module, which uses data-driven optimization algorithms and integrates road scenario simulation technology to dynamically restore complex traffic test conditions; The authenticity assessment and index quantification module is used to measure the authenticity of simulation results using a dictionary-based evaluation system and quantify the velocity distribution, density distribution and trajectory error indicators. The virtual test report generation module is used to feed back simulation data to the test platform, optimize model parameters and generate virtual test reports.

9. The data-driven urban traffic simulation model evaluation system according to claim 8, characterized in that: The data-driven approach involves obtaining the real-world speed and position of the intelligent target and mapping it onto high-definition street scenes to simulate complex traffic test conditions.

10. The data-driven urban traffic simulation model evaluation system according to claim 9, characterized in that: The intelligent object detection and real-time tracking module further includes: Visual feature extraction module, used to identify road scenes and extract key visual information; The target recognition and tracking module is used to pre-train the YOLOv8 network using a labeled dataset. The pre-trained YOLOv8 network is then used to extract key road feature information to identify intelligent targets. The module then outputs the center coordinates (x, y) of the bounding box and the category label, providing a spatial position reference for subsequent motion analysis.

Citation Information

Cited By

  • Experimental method based on crowd disaster video data and virtual scene reconstruction

    CN121527327A