Dynamic traffic flow sensing and congestion warning method and system based on urban traffic patrol robot

Abstract

The application provides a dynamic traffic flow perception and congestion early warning method and system based on a city traffic inspection robot, relates to the technical field of computer vision and roadside intelligent perception device control, and the method comprises the following steps: through the synergistic effect of three core technologies, namely, continuous traffic flow parameter collection along a road section, accurate vehicle motion state estimation based on a laser radar and congestion early warning based on microscopic traffic behavior analysis, real-time perception of road section level traffic flow and early warning of congestion are realized, a continuous traffic flow image of the whole road section can be constructed, the measurement accuracy is higher in a low-speed congestion scene, early warning of congestion is realized instead of post-confirmation, and the early warning reliability is improved through multi-dimensional feature fusion.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method and system for dynamic traffic flow perception and congestion early warning based on an urban traffic inspection robot. Background Technology

[0002] Real-time and accurate perception of traffic flow parameters, and timely and effective congestion warnings, are of great practical significance for traffic signal timing optimization, dynamic route guidance, and traffic emergency management.

[0003] Currently, urban traffic flow sensing mainly relies on the following technical solutions: (I) Induction coil detection technology. Induction coils are traditional traffic detection devices. They detect changes in the magnetic field caused by vehicles passing by by burying a loop coil under the road surface, and calculate traffic parameters such as vehicle flow, speed, time occupancy, and length. However, the coil must be directly buried in the lane during installation or maintenance, which requires cutting the road surface and will temporarily obstruct traffic; the cuts for burying the coil will soften the road surface and damage the road structure, especially at signalized intersections; the coil is susceptible to the effects of natural environments such as freezing and roadbed subsidence; the detection accuracy decreases significantly when traffic is congested and the distance between vehicles is less than 3 meters; in addition, the coil is easily damaged by heavy vehicles, making repair difficult and costly; (II) Geomagnetic detection technology

[0004] Geomagnetic detectors use high-sensitivity sensors to detect changes in the magnetic field of parking spaces or road sections to determine the presence of vehicles. They can collect data such as vehicle flow, occupancy rate, and driving direction. However, a single detector is difficult to accurately measure vehicle speed, and the vehicle type classification error is relatively large. The detection accuracy is affected by many factors such as the quality of the sensor, the reliability of the algorithm, and the stability of communication. During peak traffic hours, when vehicles are closely following each other, it is difficult to accurately determine the actual number of vehicles. (III) Microwave / millimeter-wave radar detection technology: Microwave vehicle detectors use radar wave detection technology to detect traffic data such as traffic flow, average vehicle speed, vehicle type, and lane occupancy rate in real time. However, due to the limitations of the detection principle, the radar cannot receive reflected electromagnetic waves from all parts of the target, so it cannot accurately measure the target outline and size, and it is difficult to accurately distinguish the target type. In congested traffic and road sections with many large vehicles and uneven vehicle type distribution, the measurement accuracy will be greatly affected due to obstruction. In addition, the accuracy of vehicle type classification is not high, and misjudgment is relatively serious.

[0005] (iv) Video detection technology: Video vehicle detectors acquire road images through video cameras, set up virtual loops within the video range, and detect vehicle flow and speed through image processing and recognition technology. However, this method has the following significant limitations: First, fixed-point perception blind spots. High-position video cameras are installed at fixed points and can only monitor local traffic conditions within their coverage area. There are perception blind spots between two monitoring points, making it impossible to obtain continuous traffic flow information at the road segment level and difficult to construct a continuous traffic flow profile for the entire road segment. Second, vehicle occlusion leads to inaccurate counting. When vehicles are densely queued, high-position videos suffer from severe mutual occlusion, especially at intersections where cameras are installed close to the edge of one side of the road and have limited height, resulting in more severe occlusion. Existing methods struggle to accurately count the number of queued vehicles and measure queue length, leading to significant deviations in traffic flow parameter estimation.

[0006] Third, speed estimation errors are large in low-speed scenes. The overhead view angle of high-angle video limits the accuracy of vehicle speed estimation. Speed ​​estimation methods based on image optical flow suffer from drastically increased errors in congested scenarios where vehicles are moving at low speeds or alternating between stopping and starting, due to the extremely small range of vehicle movement.

[0007] Fourth, poor environmental adaptability. Video detectors are easily affected by changes in lighting and weather conditions, under which their performance degrades sharply; detection accuracy also decreases significantly in complex environments such as vehicle obstruction and shadows. Fifth, lacking the ability to perceive microscopic traffic behavior. V. Multi-sensor fusion solution

[0008] In recent years, multi-source sensor fusion technology has gradually attracted attention. By fusing millimeter-wave radar, visual sensors, and lidar, they complement each other in terms of physical characteristics. However, most existing fusion solutions are still based on fixed-point deployment, which essentially does not escape the limitations of fixed monitoring modes and still suffers from the problem of perception blind spots. At the same time, roadside sensors are far from vehicles and also face the problems of obstruction and reduced accuracy in congested scenarios. Summary of the Invention

[0009] The purpose of this application is to provide a dynamic traffic flow perception and congestion early warning method and system based on urban traffic inspection robots, in order to solve the problems of existing urban traffic flow perception and congestion early warning technologies, which have varying degrees of blind spots, occlusion interference, insufficient accuracy of low-speed measurement, poor environmental adaptability, and lack of micro-behavioral perception.

[0010] In view of the above problems, this application provides a method and system for dynamic traffic flow perception and congestion early warning based on urban traffic inspection robots.

[0011] The first aspect of this application provides a method for dynamic traffic flow perception and congestion early warning based on an urban traffic inspection robot, which collects multimodal data at a predetermined collection frequency. The multimodal data includes three-dimensional point cloud data, image data, positioning data, and the driving speed of the inspection robot itself.

[0012] Based on the three-dimensional point cloud data, vehicles around the inspection robot are detected and the motion state parameters of each detected vehicle are estimated. The motion state parameters include the three-dimensional position, velocity and acceleration of each vehicle.

[0013] Based on the detected motion state parameters of each vehicle, traffic flow parameters at the road segment level are calculated.

[0014] Based on the detected motion state parameters of each vehicle, micro-level traffic behavior indicators are calculated.

[0015] The traffic flow parameters and the micro-traffic behavior indicators are combined into a traffic state feature vector, and the traffic state in future time periods is predicted based on the traffic state feature vector.

[0016] Furthermore, the step of detecting vehicles around the inspection robot and estimating the motion state parameters of each detected vehicle includes:

[0017] The calculation is performed using the formulas V_i(t) = {center_i(t), size_i(t), heading_i(t), v_i(t), a_i(t)}, v_i(t) = (center_i(t) - center_i(t-Δt)) / Δt, and a_i(t) = (v_i(t) - v_i(t-Δt)) / Δt. Here, V_i(t) represents the state information vector of the i-th vehicle detected at time t; center_i(t) represents the coordinates of the center point of the i-th vehicle in the three-dimensional world coordinate system; size_i(t) represents the geometric dimensions of the three-dimensional bounding box of the i-th vehicle; heading_i(t) represents the heading angle of the i-th vehicle; v_i(t) represents the instantaneous velocity of the i-th vehicle; a_i(t) represents the instantaneous acceleration of the i-th vehicle; and Δt represents the time interval between two consecutive frames.

[0018] Furthermore, the step of calculating the segment-level traffic flow parameters based on the detected motion state parameters of each vehicle includes:

[0019] The road segment is evenly divided into several sub-segments according to a predetermined length. The traffic flow parameters are calculated for each sub-segment. The formula for calculating the traffic volume Q_seg is Q_seg = N_pass / T_obs.

[0020] Wherein, N_pass represents the total number of vehicles passing through the cross section of this sub-segment within the preset observation time T_obs;

[0021] The average vehicle speed V_avg is calculated as V_avg = Σ_{i=1}^{N_v} ||v_i|| / N_v, where N_v represents the total number of vehicles detected in the current sub-segment, and v_i represents the speed vector of the i-th vehicle. The lane occupancy rate O_lane is calculated as O_lane = Σ_{i=1}^{N_v} l_i / L_seg, where l_i represents the body length of the i-th vehicle, and L_seg represents the segment division length.

[0022] The queue length L_queue is estimated by detecting the spatial range of continuously slow-moving vehicles: L_queue = max(s_i) - min(s_i), where ||v_i|| < v_queue, s_i represents the position coordinates of the i-th slow-moving vehicle along the road segment direction, and v_queue is the low-speed determination threshold.

[0023] Furthermore, the step of calculating micro-traffic behavior indicators based on the detected motion state parameters of each vehicle includes:

[0024] According to the formula BF = Σ_{i=1}^{N_v} Σ_{t} I(a_i(t) < -a_brake) / (N_v · T_obs), CV_gap = σ_gap / μ_gap, SFI = (1 / N_v) · Σ_{i=1}^{N_v} σ_v_i / μ_v_i is calculated, where I(·) is an indicator function, which takes the value of 1 when the condition in parentheses is met, and takes the value of 0 otherwise. a_i(t) represents the instantaneous acceleration value of the i-th vehicle at time t, a_brake is the braking acceleration judgment threshold, N_v represents the total number of vehicles in the observation range, T_obs represents the observation time length, σ_gap represents the standard deviation of the distance between all adjacent vehicles in the observation segment, μ_gap represents the arithmetic mean of the distance between all adjacent vehicles, σ_v_i represents the standard deviation of the speed of the i-th vehicle in the time dimension, and μ_v_i represents the mean of the speed of the i-th vehicle in the time dimension.

[0025] Furthermore, the step of combining the traffic flow parameters and the micro-traffic behavior indicators into a traffic state feature vector, and predicting the traffic state for future periods based on the traffic state feature vector, includes:

[0026] The traffic state feature vector X(t, s) is expressed as: X(t, s) = [Q_seg, V_avg, O_lane, L_queue, BF, CV_gap, SFI]^T, where Q_seg is the traffic flow, V_avg is the average vehicle speed, O_lane is the lane occupancy rate, L_queue is the queue length, BF is the braking frequency, CV_gap is the coefficient of variation of vehicle spacing, and SFI is the speed fluctuation index;

[0027] The traffic state for future time periods is predicted using the formula X_pred(t+T_p, s) = STTransformer({X(t-T_h, s), ..., X(t, s)}), where X_pred(t+T_p, s) represents the traffic state feature vector at location s on the road segment at time t+T_p in the future; STTransformer(·) represents the spatiotemporal Transformer network model; {X(t-T_h,s), ..., X(t, s)} represents the continuous feature vector sequence within the historical observation window; T_h represents the length of the historical observation window; and T_p represents the prediction time span.

[0028] A second aspect of this application provides a dynamic traffic flow perception and congestion early warning system based on an urban traffic inspection robot, characterized in that it includes:

[0029] The acquisition module is used to acquire multimodal data at a predetermined acquisition frequency. The multimodal data includes three-dimensional point cloud data, image data, positioning data, and the travel speed of the inspection robot itself.

[0030] The detection module is used to detect vehicles around the inspection robot based on the three-dimensional point cloud data and estimate the motion state parameters of each detected vehicle, including the three-dimensional position, velocity and acceleration of each vehicle.

[0031] The calculation module is used to calculate the traffic flow parameters at the road segment level based on the motion state parameters of each detected vehicle.

[0032] The calculation module is also used to calculate micro-traffic behavior indicators based on the detected motion state parameters of each vehicle.

[0033] The prediction module is used to combine the traffic flow parameters and the micro-traffic behavior indicators into a traffic state feature vector, and predict the traffic state for future periods based on the traffic state feature vector.

[0034] Further, the detection module is specifically used to calculate according to the formulas V_i(t) = {center_i(t), size_i(t), heading_i(t), v_i(t), a_i(t)}, v_i(t) = (center_i(t) - center_i(t-Δt)) / Δt, and a_i(t) = (v_i(t) - v_i(t-Δt)) / Δt, where V_i(t) represents the state information vector of the i-th vehicle detected at time t; center_i(t) represents the coordinates of the center point of the i-th vehicle in the three-dimensional world coordinate system; size_i(t) represents the geometric dimensions of the three-dimensional bounding box of the i-th vehicle; heading_i(t) represents the heading angle of the i-th vehicle; v_i(t) represents the instantaneous velocity of the i-th vehicle; a_i(t) represents the instantaneous acceleration of the i-th vehicle; and Δt represents the time interval between two consecutive frames.

[0035] Further, the calculation module is specifically used to uniformly divide the road segment into several sub-segments according to a predetermined length, and calculate the traffic flow parameters for each sub-segment. The calculation formula for the traffic flow Q_seg is Q_seg = N_pass / T_obs, where N_pass represents the total number of vehicles passing through the cross-section of the sub-segment within the preset observation time T_obs; the calculation formula for the average vehicle speed V_avg is V_avg = Σ_{i=1}^{N_v} ||v_i|| / N_v, where N_v represents the total number of vehicles detected in the current sub-segment, v_i represents the speed vector of the i-th vehicle; the calculation formula for the lane occupancy rate O_lane is O_lane = Σ_{i=1}^{N_v} l_i / L_seg, where l_i represents the body length of the i-th vehicle, and L_seg represents the division length of the sub-segment; the queue length L_queue is estimated by detecting the spatial range of continuously slow-moving vehicles: L_queue = max(s_i) - min(s_i), where ||v_i|| < v_queue, s_i represents the position coordinates of the i-th low-speed vehicle along the road segment direction, and v_queue is the low-speed judgment threshold.

[0036] Further, the calculation module is specifically used to calculate BF = Σ_{i=1}^{N_v} Σ_{t} I(a_i(t) < -a_brake) / (N_v · T_obs), CV_gap = σ_gap / μ_gap, and SFI = (1 / N_v) · Σ_{i=1}^{N_v} σ_v_i / μ_v_i is calculated, where I(·) is an indicator function, which takes the value of 1 when the condition in parentheses is met, and takes the value of 0 otherwise. a_i(t) represents the instantaneous acceleration value of the i-th vehicle at time t, a_brake is the braking acceleration judgment threshold, N_v represents the total number of vehicles in the observation range, T_obs represents the observation time length, σ_gap represents the standard deviation of the distance between all adjacent vehicles in the observation segment, μ_gap represents the arithmetic mean of the distance between all adjacent vehicles, σ_v_i represents the standard deviation of the speed of the i-th vehicle in the time dimension, and μ_v_i represents the mean of the speed of the i-th vehicle in the time dimension.

[0037] Further, the prediction module is specifically used to represent the traffic state feature vector X(t, s) as: X(t, s) = [Q_seg, V_avg, O_lane, L_queue, BF, CV_gap, SFI]^T, where Q_seg is the traffic flow, V_avg is the average vehicle speed, O_lane is the lane occupancy rate, L_queue is the queue length, BF is the braking frequency, CV_gap is the coefficient of variation of vehicle spacing, and SFI is the speed fluctuation index; according to the formula X_pred(t+T_p, s) = STTransformer({X(t-T_h, s), ..., X(t, s)}), the traffic state in the future time period is predicted, where X_pred(t+T_p, s) represents the predicted traffic state feature vector at the road segment location s at the future time t+T_p; STTransformer(·) represents the spatiotemporal Transformer network model, {X(t-T_h, s), ..., X(t, s)} represents a continuous sequence of feature vectors within a historical observation window; T_h represents the length of the historical observation window; and T_p represents the prediction time span.

[0038] One or more technical solutions provided in this application have at least the following technical effects or advantages: The dynamic traffic flow perception and congestion early warning method based on urban traffic inspection robots proposed in this invention achieves real-time perception of traffic flow at the road segment level and early warning of congestion through the synergistic effect of three core technologies: continuous traffic flow parameter collection along the road segment, accurate vehicle motion state estimation based on lidar, and congestion early warning based on microscopic traffic behavior analysis. Compared with the traditional traffic flow perception scheme based on high-position video, the beneficial effects of this invention are mainly reflected in the following four aspects: (1) It eliminates the perception blind spots of fixed points. The inspection robot continuously travels along the road segment to collect data, and there are no monitoring blank areas between fixed cameras, which can construct a continuous traffic flow profile of the entire road segment; (2) It has higher measurement accuracy in low-speed congestion scenarios. The lidar speed estimation method based on point cloud registration can still maintain centimeter-level accuracy under low-speed conditions, which is significantly better than the high-position video method based on image optical flow; (3) It realizes early warning of congestion rather than post-event confirmation. By capturing early warning signals of congestion through micro-traffic behavior indicators and combining them with a spatiotemporal Transformer network to predict the future period, valuable advance response time is provided for traffic management; (4) Multi-dimensional feature fusion improves the reliability of early warning. A 7-dimensional feature vector is constructed by combining macro-traffic flow parameters and micro-traffic behavior indicators to comprehensively judge the risk of congestion from multiple dimensions, effectively reducing the false alarms and false alarms that may be caused by a single indicator.

[0039] The above description is merely an overview of the technical solution of this application. To better understand the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0041] Figure 1 This is a flowchart illustrating the dynamic traffic flow perception and congestion early warning method based on urban traffic inspection robots provided in this application.

[0042] Figure 2This is a schematic diagram of the dynamic traffic flow perception and congestion early warning system based on urban traffic inspection robots provided in this application. Detailed Implementation

[0043] This application provides a method and system for dynamic traffic flow perception and congestion early warning based on urban traffic inspection robots. It addresses the technical problems in existing technologies where robot task scheduling and energy management are independent, leading to high cluster energy consumption and low task execution efficiency. The method achieves the technical effect of improving cluster task scheduling efficiency and energy utilization efficiency by optimizing the coupling of robot cluster energy and tasks, enabling real-time generation of energy storage scheduling strategies.

[0044] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. It should also be noted that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, not all of them.

[0045] Example 1, as Figure 1 As shown, this application provides a method for dynamic traffic flow perception and congestion early warning based on urban traffic inspection robots. The method includes:

[0046] 101. Collect multimodal data at a predetermined collection frequency. The multimodal data includes three-dimensional point cloud data, image data, positioning data, and the travel speed of the inspection robot itself.

[0047] Specifically, the inspection robot is equipped with a 360-degree LiDAR, forward-looking and rear-looking visible light cameras, a GPS / RTK positioning module, and an odometer. It continuously travels along urban road sections and collects multimodal data in real time. The set of multimodal data collected by the inspection robot at time t and position s is defined as follows:

[0048] D(t, s) = {P_lidar(t), I_front(t), I_rear(t), GPS(t), v_ego(t)}, where D(t,s) represents the complete data set collected at location s on the road at time t; P_lidar(t) represents the 3D point cloud data collected by the 360-degree lidar at time t, including distance and reflection intensity information of the surrounding environment; I_front(t) represents the RGB image collected by the forward-looking visible light camera; I_rear(t) represents the RGB image collected by the rear-looking visible light camera; GPS(t) represents the centimeter-level positioning coordinates output by the RTK-GPS module, including longitude, latitude, and elevation information; and v_ego(t) represents the inspection robot's own driving speed, obtained by integrating the wheel speed encoding from the odometer. The inspection robot continuously collects the above multimodal data at a fixed frequency of 10Hz, forming a spatiotemporally continuous data sequence as it travels along the road segment, ensuring no blind spots.

[0049] 102. Based on the three-dimensional point cloud data, detect vehicles around the inspection robot and estimate the motion state parameters of each detected vehicle.

[0050] Specifically, the calculation is performed according to the formulas V_i(t) = {center_i(t), size_i(t), heading_i(t), v_i(t), a_i(t)}, v_i(t) = (center_i(t) - center_i(t-Δt)) / Δt, and a_i(t) = (v_i(t)- v_i(t-Δt)) / Δt. Here, V_i(t) represents the state information vector of the i-th vehicle detected at time t; center_i(t) represents the coordinates of the center point of the i-th vehicle in the three-dimensional world coordinate system; size_i(t) represents the geometric dimensions of the three-dimensional bounding box of the i-th vehicle; heading_i(t) represents the heading angle of the i-th vehicle; v_i(t) represents the instantaneous velocity of the i-th vehicle; a_i(t) represents the instantaneous acceleration of the i-th vehicle; and Δt represents the time interval between two consecutive frames.

[0051] For example, 360-degree LiDAR point clouds can be used to accurately detect and 3D locate vehicles around an inspection robot. This invention employs CenterPoint, a point cloud-based 3D target detection network, as the detection backbone network. This network can efficiently process sparse and unordered point cloud data and output the 3D detection results of vehicles in real time. For each detected vehicle, its 3D bounding box and motion state parameters are output: V_i(t) = {center_i(t), size_i(t), heading_i(t), v_i(t), a_i(t)}

[0052] Wherein, V_i(t) represents the complete state information vector of the i-th vehicle detected at time t; center_i(t) = (x_i, y_i, z_i) represents the coordinates of the center point of the i-th vehicle in the three-dimensional world coordinate system, where x_i and y_i are the horizontal position coordinates and z_i is the height coordinate; size_i(t) = (l_i, w_i, h_i) represents the geometric dimensions of the three-dimensional bounding box of the i-th vehicle, where l_i is the vehicle length, w_i is the vehicle width, and h_i is the vehicle height; heading_i(t) represents the heading angle of the i-th vehicle, with a value range of [0, 2π); v_i(t) represents the instantaneous velocity scalar value of the i-th vehicle; and a_i(t) represents the instantaneous acceleration scalar value of the i-th vehicle.

[0053] Specifically, vehicle speed and acceleration are calculated by the difference between adjacent frames, using the following formula: v_i(t) = (center_i(t) - center_i(t-Δt)) / Δt

[0054] Where v_i(t) represents the instantaneous velocity vector of the i-th vehicle at the current time t; center_i(t) and center_i(t-Δt) represent the three-dimensional coordinate vectors of the center of the i-th vehicle in the current frame and the previous frame (i.e., at time t-Δt), respectively; Δt represents the time interval between two consecutive frames, which is a fixed value of 0.1 seconds, corresponding to a data acquisition frequency of 10Hz. a_i(t) = (v_i(t) - v_i(t-Δt)) / Δt, where a_i(t) represents the instantaneous acceleration vector of the i-th vehicle at the current time t; v_i(t) and v_i(t-Δt) represent the instantaneous velocity vectors of the i-th vehicle in the current frame and the previous frame, respectively; Δt is also a time interval of 0.1 seconds. It should be noted that the above velocity and acceleration calculations need to compensate for the motion of the inspection robot itself, that is, the vehicle's motion state is transformed from the following robot coordinate system to the stationary world coordinate system through coordinate transformation to obtain the vehicle's true ground motion parameters.

[0055] The lidar system on the inspection robot offers significant technological advantages over high-position video in estimating vehicle motion. LiDAR directly measures the three-dimensional distance information on the vehicle surface, unaffected by changes in lighting or shadows, and maintains stable measurement accuracy even at night and in adverse weather conditions. Particularly in low-speed, congested scenarios, the velocity estimation method based on image optical flow in high-position video suffers a sharp increase in error due to the extremely small amplitude of vehicle movement, while the velocity estimation method based on point cloud registration in lidar can still maintain centimeter-level displacement measurement accuracy, thus enabling precise perception of the motion state of low-speed vehicles.

[0056] 103. Based on the detected motion state parameters of each vehicle, calculate the traffic flow parameters at the road segment level.

[0057] Specifically, the road segment is uniformly divided into several sub-segments according to a predetermined length. The traffic flow parameters are calculated for each sub-segment. The traffic volume Q_seg is calculated as Q_seg = N_pass / T_obs, where N_pass represents the total number of vehicles passing through the cross-section of that sub-segment within a preset observation time T_obs. The average vehicle speed V_avg is calculated as V_avg = Σ_{i=1}^{N_v} ||v_i|| / N_v, where N_v represents the total number of vehicles detected in the current sub-segment, and v_i represents the speed vector of the i-th vehicle. The lane occupancy rate O_lane is calculated as O_lane = Σ_{i=1}^{N_v}l_i / L_seg, where l_i represents the vehicle length of the i-th vehicle, and L_seg represents the segment division length. The queue length L_queue is estimated by detecting the spatial range of continuously slow-moving vehicles: L_queue = max(s_i) - min(s_i), where ||v_i|| < v_queue, s_i represents the position coordinates of the i-th low-speed vehicle along the road segment direction, and v_queue is the low-speed judgment threshold.

[0058] For example, based on vehicle detection and motion status data continuously collected along the road segment, road segment-level traffic flow parameters are calculated. The road segment is uniformly divided into several sub-segments of a fixed length L_seg (set to 50 meters), and the following four core traffic flow parameters are calculated for each sub-segment: First, the traffic volume Q_seg, defined as the number of vehicles passing through the sub-segment per unit time: Q_seg = N_pass / T_obs.

[0059] Where Q_seg represents the traffic flow index of the sub-segment, in vehicles per hour; N_pass represents the total number of vehicles passing through the cross-section of the sub-segment within the preset observation time T_obs; T_obs represents the observation time length, in hours. Traffic flow is a fundamental parameter for measuring road capacity and traffic load.

[0060] Second, the average vehicle speed V_avg is defined as the arithmetic mean of the speeds of all detected vehicles within the sub-segment: V_avg = Σ_{i=1}^{N_v} ||v_i|| / N_v.

[0061] Where V_avg represents the average speed of all vehicles in the sub-segment, in km / h or m / s; N_v represents the total number of vehicles detected in the current sub-segment; v_i represents the velocity vector of the i-th vehicle, and ||v_i|| represents the L2 norm of the velocity vector of the i-th vehicle, i.e., the scalar value of the velocity; Σ_{i=1}^{N_v} (·) represents the summation of the calculation results for vehicles from the 1st to the N_vth vehicle in the sub-segment. Average vehicle speed is a core indicator for evaluating the traffic efficiency of a road segment.

[0062] Third, lane occupancy rate O_lane is defined as the proportion of space occupied by vehicles within a sub-segment:

[0063] O_lane = Σ_{i=1}^{N_v} l_i / L_seg.

[0064] Where O_lane represents lane occupancy, a dimensionless ratio ranging from [0, 1]; l_i represents the length of the i-th vehicle; L_seg represents the segment division length, fixed at 50 meters; Σ_{i=1}^{N_v} l_i represents the sum of the lengths of all vehicles in the current segment. Lane occupancy directly reflects the utilization of road space; when the occupancy exceeds 0.3, it usually means that traffic is beginning to approach saturation.

[0065] Fourth, the queue length L_queue is estimated by detecting the spatial range of continuously moving low-speed vehicles: L_queue = max(s_i) - min(s_i), where ||v_i|| < v_queue.

[0066] Where L_queue represents the queue length in meters; s_i represents the position coordinates of the i-th slow-moving vehicle along the road segment; max(s_i) and min(s_i) represent the position coordinates of the front and rearmost vehicles among all slow-moving vehicles, respectively; v_queue is the slow-moving threshold, set to 5 km / h, vehicles below this speed are considered to be in or near a queue; ||v_i|| < v_queue indicates that queue analysis is only performed on vehicles with speeds below 5 km / h. Queue length is the most intuitive indicator of congestion severity.

[0067] 104. Calculate micro-traffic behavior indicators based on the detected motion state parameters of each vehicle.

[0068] Specifically, according to the formula BF = Σ_{i=1}^{N_v} Σ_{t} I(a_i(t) < -a_brake) / (N_v·T_obs), CV_gap = σ_gap / μ_gap, SFI = (1 / N_v) · Σ_{i=1}^{N_v} σ_v_i / μ_v_i is calculated, where I(·) is an indicator function, which takes the value of 1 when the condition in parentheses is met, and takes the value of 0 otherwise. a_i(t) represents the instantaneous acceleration value of the i-th vehicle at time t, a_brake is the braking acceleration judgment threshold, N_v represents the total number of vehicles in the observation range, T_obs represents the observation time length, σ_gap represents the standard deviation of the distance between all adjacent vehicles in the observation segment, μ_gap represents the arithmetic mean of the distance between all adjacent vehicles, σ_v_i represents the standard deviation of the speed of the i-th vehicle in the time dimension, and μ_v_i represents the mean of the speed of the i-th vehicle in the time dimension.

[0069] For example, this invention innovatively utilizes the unique low-angle advantage of the inspection robot situated within traffic flow to capture early warning signals of congestion by closely sensing micro-traffic behavior, thus achieving proactive early warning to "prevent problems before they occur." The following three categories of micro-traffic behavior indicators are defined:

[0070] (1) Braking Frequency (BF) is defined as the number of times surrounding vehicles brake significantly per unit time:

[0071] BF = Σ_{i=1}^{N_v} Σ_{t} I(a_i(t) < -a_brake) / (N_v · T_obs)

[0072] Wherein, BF represents the braking frequency index, measured in times per (vehicle·hour); I(·) is an indicator function, taking a value of 1 when the condition within the parentheses is met, and 0 otherwise; a_i(t) represents the instantaneous acceleration value of the i-th vehicle at time t; a_brake is the braking acceleration judgment threshold, set to -2 m / s², meaning that a significant braking event is considered to have occurred when the vehicle deceleration exceeds 2 m / s²; N_v represents the total number of vehicles within the observation range; T_obs represents the observation time length; Σ_{i=1}^{N_v} Σ_{t} (·) represents the double summation of all vehicles at all time points. A higher BF value indicates that the driver frequently performs sudden deceleration operations, which is an important early warning signal of unstable traffic flow ahead.

[0073] (2) The coefficient of variation of vehicle spacing CV_gap is defined as the ratio of the standard deviation of the spacing between adjacent vehicles to the mean: CV_gap = σ_gap / μ_gap;

[0074] Here, CV_gap represents the coefficient of variation of vehicle spacing, a dimensionless indicator; σ_gap represents the standard deviation of the distance between all adjacent vehicles within the observed road segment, reflecting the dispersion of spacing; and μ_gap represents the arithmetic mean of the distances between all adjacent vehicles. A larger CV_gap indicates a more uneven distribution of vehicle spacing—with denser areas and sparser areas, a typical characteristic of congestion forming or dissipating. Normally, CV_gap is less than 0.3, while a value exceeding 0.5 indicates that traffic flow has begun to become significantly unstable.

[0075] (3) Speed ​​fluctuation index SFI (, defined as the coefficient of variation of vehicle speed over time):

[0076] SFI = (1 / N_v) · Σ_{i=1}^{N_v} σ_v_i / μ_v_i;

[0077] Here, SFI stands for Speed ​​Fluctuation Index, a dimensionless indicator; N_v represents the total number of observed vehicles; σ_v_i represents the standard deviation of the speed of the i-th vehicle over time, reflecting the degree of speed fluctuation over time; and μ_v_i represents the mean speed of the i-th vehicle over time. A larger SFI indicates more drastic speed fluctuations—sometimes moving quickly, sometimes suddenly decelerating or even stopping—a typical quantitative characteristic of the "stop-and-go" congestion pattern. In free-flow conditions, SFI is typically below 0.1; when it exceeds 0.3, it indicates that the traffic flow has entered a clear stop-and-go state.

[0078] 105. Combine the traffic flow parameters and the micro-traffic behavior indicators into a traffic state feature vector, and predict the traffic state for future periods based on the traffic state feature vector.

[0079] Specifically, the traffic state feature vector X(t, s) is expressed as: X(t, s) = [Q_seg, V_avg, O_lane, L_queue, BF, CV_gap, SFI]^T, where Q_seg is the traffic flow, V_avg is the average vehicle speed, O_lane is the lane occupancy rate, L_queue is the queue length, BF is the braking frequency, CV_gap is the coefficient of variation of vehicle spacing, and SFI is the speed fluctuation index; the traffic state in the future is predicted according to the formula X_pred(t+T_p, s) = STTransformer({X(t-T_h, s), ..., X(t, s)}), where X_pred(t+T_p, s) represents the predicted traffic state feature vector at location s on the road segment at time t+T_p in the future; STTransformer(·) represents the spatiotemporal Transformer network model, {X(t-T_h,s), ..., X(t, s)} represents a continuous sequence of feature vectors within the historical observation window; T_h represents the length of the historical observation window; T_p represents the prediction time span.

[0080] For example, this invention designs a congestion early warning model based on spatiotemporal feature fusion, which integrates macroscopic traffic flow parameters and microscopic traffic behavior indicators to predict congestion. The segment-level traffic flow parameters and microscopic behavior indicators are combined into a multidimensional feature vector: X(t, s) = [Q_seg, V_avg, O_lane, L_queue, BF, CV_gap, SFI]^T;

[0081] Where X(t, s) represents the traffic state feature vector at time t and location s on road segment, which is a 7-dimensional column vector; Q_seg is the traffic flow; V_avg is the average vehicle speed; O_lane is the lane occupancy rate; L_queue is the queue length; BF is the braking frequency index; CV_gap is the coefficient of variation of vehicle spacing; and SFI is the speed fluctuation index. The superscript T indicates the vector transpose operation, converting the row vector into a column vector for convenient subsequent matrix operations.

[0082] Specifically, a temporal prediction network is used to predict traffic conditions for future periods. This invention employs a spatiotemporal Transformer network, which can simultaneously capture the dependencies of traffic flow data in both the temporal and spatial dimensions. The feature vector sequence continuously collected along the road segment is used as the input of historical observations to predict the traffic conditions for future periods: X_pred(t+T_p, s) = STTransformer({X(t-T_h, s), ..., X(t, s)});

[0083] Where X_pred(t+T_p, s) represents the traffic state feature vector at location s of the road segment at the predicted future time t+T_p; STTransformer(·) represents the spatiotemporal Transformer network model, which models spatiotemporal dependencies simultaneously through a self-attention mechanism; {X(t-T_h, s), ..., X(t, s)} represents the continuous feature vector sequence within the historical observation window; T_h represents the length of the historical observation window, set to 10 minutes; T_p represents the prediction time span, set to 5 to 15 minutes, which can be dynamically selected according to traffic management needs.

[0084] Specifically, based on the predicted future traffic state feature vector, the probability of congestion is calculated through a fully connected layer and a Sigmoid activation function: P_congestion = σ(W · X_pred + b);

[0085] Wherein, P_congestion represents the probability of congestion occurring on this road segment in the future time period, with a value range of [0, 1]; σ(·) represents the Sigmoid activation function, whose mathematical expression is σ(z) = 1 / (1+e^{-z}), which maps any real number to a probability value between 0 and 1; W represents the learnable weight vector of the fully connected layer, with its dimension matching the feature vector X_pred; b represents the bias term, which is a learnable scalar parameter; (·) represents the vector dot product operation. When P_congestion exceeds the warning threshold θ_warn (set to 0.7), the system automatically triggers a congestion warning and pushes a warning notification containing information such as location, time, and expected severity to the traffic management center.

[0086] Furthermore, a loss function is constructed for model training, and model optimization and parameter updates are performed. This invention employs a multi-task joint training strategy, and the model loss function consists of three parts: vehicle detection loss L_det, traffic flow parameter estimation loss L_flow, and congestion prediction loss L_pred.

[0087] L_total = L_det + λ_1 · L_flow + λ_2 · L_pred

[0088] Where L_total represents the total training loss of the model; L_det represents the loss of the 3D vehicle detection network, specifically the sum of Focal Loss and Smooth-L1 loss of bounding box regression, used to optimize vehicle detection accuracy; L_flow represents the mean squared error (MSE) loss of traffic flow parameter estimation, used to optimize the estimation accuracy of parameters such as traffic flow and average vehicle speed; L_pred represents the binary cross-entropy loss of congestion prediction, used to optimize the accuracy of congestion classification; λ_1 and λ_2 represent the weight coefficients of traffic flow loss and prediction loss relative to detection loss, respectively, set to 0.5 and 1.0, used to balance the contribution ratio of each subtask in joint training.

[0089] It should be understood that those skilled in the art can make improvements and modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

[0090] Example 2 is based on the same inventive concept as the dynamic traffic flow perception and congestion early warning method based on urban traffic inspection robots in the previous examples, such as... Figure 2 As shown, this application provides a dynamic traffic flow perception and congestion early warning system based on an urban traffic inspection robot, wherein the dynamic traffic flow perception and congestion early warning system based on an urban traffic inspection robot includes:

[0091] The acquisition module 21 is used to acquire multimodal data at a predetermined acquisition frequency. The multimodal data includes three-dimensional point cloud data, image data, positioning data, and the travel speed of the inspection robot itself.

[0092] The detection module 22 is used to detect vehicles around the inspection robot based on the three-dimensional point cloud data and estimate the motion state parameters of each detected vehicle. The motion state parameters include the three-dimensional position, velocity and acceleration of each vehicle.

[0093] Calculation module 23 is used to calculate road segment-level traffic flow parameters based on the detected motion state parameters of each vehicle;

[0094] The calculation module 23 is also used to calculate micro-traffic behavior indicators based on the detected motion state parameters of each vehicle.

[0095] The prediction module 24 is used to combine the traffic flow parameters and the micro traffic behavior indicators into a traffic state feature vector, and predict the traffic state in future time periods based on the traffic state feature vector.

[0096] Further, the detection module 22 is specifically used to calculate according to the formulas V_i(t) = {center_i(t), size_i(t), heading_i(t), v_i(t), a_i(t)}, v_i(t) = (center_i(t) - center_i(t-Δt)) / Δt, and a_i(t) = (v_i(t) - v_i(t-Δt)) / Δt, where V_i(t) represents the state information vector of the i-th vehicle detected at time t; center_i(t) represents the coordinates of the center point of the i-th vehicle in the three-dimensional world coordinate system; size_i(t) represents the geometric dimensions of the three-dimensional bounding box of the i-th vehicle; heading_i(t) represents the heading angle of the i-th vehicle; v_i(t) represents the instantaneous velocity of the i-th vehicle; a_i(t) represents the instantaneous acceleration of the i-th vehicle; and Δt represents the time interval between two consecutive frames.

[0097] Further, the calculation module 23 is specifically used to uniformly divide the road segment into several sub-segments according to a predetermined length, and calculate the traffic flow parameters for each sub-segment. The calculation formula for the traffic flow Q_seg is Q_seg = N_pass / T_obs, where N_pass represents the total number of vehicles passing through the cross-section of the sub-segment within the preset observation time T_obs; the calculation formula for the average vehicle speed V_avg is V_avg = Σ_{i=1}^{N_v} ||v_i|| / N_v, where N_v represents the total number of vehicles detected in the current sub-segment, v_i represents the speed vector of the i-th vehicle; the calculation formula for the lane occupancy rate O_lane is O_lane = Σ_{i=1}^{N_v} l_i / L_seg, where l_i represents the body length of the i-th vehicle, and L_seg represents the division length of the sub-segment; the queue length L_queue is estimated by detecting the spatial range of continuously slow-moving vehicles: L_queue = max(s_i) - min(s_i), where ||v_i|| < v_queue, s_i represents the position coordinates of the i-th low-speed vehicle along the road segment direction, and v_queue is the low-speed judgment threshold.

[0098] Further, the calculation module 23 is specifically used to calculate BF = Σ_{i=1}^{N_v} Σ_{t} I(a_i(t) < -a_brake) / (N_v · T_obs), CV_gap = σ_gap / μ_gap, SFI = (1 / N_v) ·Σ_{i=1}^{N_v} σ_v_i / μ_v_i is calculated, where I(·) is an indicator function, which takes the value of 1 when the condition in parentheses is met, and takes the value of 0 otherwise. a_i(t) represents the instantaneous acceleration value of the i-th vehicle at time t, a_brake is the braking acceleration judgment threshold, N_v represents the total number of vehicles in the observation range, T_obs represents the observation time length, σ_gap represents the standard deviation of the distance between all adjacent vehicles in the observation segment, μ_gap represents the arithmetic mean of the distance between all adjacent vehicles, σ_v_i represents the standard deviation of the speed of the i-th vehicle in the time dimension, and μ_v_i represents the mean of the speed of the i-th vehicle in the time dimension.

[0099] Further, the prediction module 24 is specifically used to represent the traffic state feature vector X(t, s) as: X(t, s) = [Q_seg, V_avg, O_lane, L_queue, BF, CV_gap, SFI]^T, where Q_seg is the traffic flow, V_avg is the average vehicle speed, O_lane is the lane occupancy rate, L_queue is the queue length, BF is the braking frequency, CV_gap is the coefficient of variation of vehicle spacing, and SFI is the speed fluctuation index; according to the formula X_pred(t+T_p, s) = STTransformer({X(t-T_h, s), ..., X(t, s)}), the traffic state in the future time period is predicted, where X_pred(t+T_p, s) represents the predicted traffic state feature vector at the road segment location s at the future time t+T_p; STTransformer(·) represents the spatiotemporal Transformer network model, {X(t-T_h, ..., X(t, s)}) = [Q_seg, V_avg, O_lane, L_queue, BF, CV_gap, SFI]^T, ..., X(t-T_h, ..., X(t, s)}) = [Q_seg, V_avg, O_lane, ..., X(t ...} represents a continuous sequence of feature vectors within a historical observation window; T_h represents the length of the historical observation window; and T_p represents the prediction time span. The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. The dynamic traffic flow perception and congestion warning method and specific examples based on urban traffic inspection robots in the aforementioned Embodiment 1 are also applicable to the dynamic traffic flow perception and congestion warning system based on urban traffic inspection robots in this embodiment. Through the foregoing detailed description of the dynamic traffic flow perception and congestion warning method based on urban traffic inspection robots, those skilled in the art can clearly understand the dynamic traffic flow perception and congestion warning system based on urban traffic inspection robots in this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0100] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0101] Obviously, those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for dynamic traffic flow perception and congestion early warning based on urban traffic inspection robots, characterized in that, include: Multimodal data is collected at a predetermined collection frequency. The multimodal data includes three-dimensional point cloud data, image data, positioning data, and the travel speed of the inspection robot itself. Based on the three-dimensional point cloud data, vehicles around the inspection robot are detected and the motion state parameters of each detected vehicle are estimated. The motion state parameters include the three-dimensional position, velocity and acceleration of each vehicle. Based on the detected motion state parameters of each vehicle, traffic flow parameters at the road segment level are calculated. Based on the detected motion state parameters of each vehicle, micro-level traffic behavior indicators are calculated. The traffic flow parameters and the micro-traffic behavior indicators are combined into a traffic state feature vector, and the traffic state in future time periods is predicted based on the traffic state feature vector.

2. The method according to claim 1, characterized in that, The step of detecting vehicles around the inspection robot and estimating the motion state parameters of each detected vehicle based on the 3D point cloud data includes: The calculation is performed using the formulas V_i(t) = {center_i(t), size_i(t), heading_i(t), v_i(t), a_i(t)}, v_i(t) = (center_i(t) - center_i(t-Δt)) / Δt, and a_i(t) = (v_i(t) - v_i(t-Δt)) / Δt. Here, V_i(t) represents the state information vector of the i-th vehicle detected at time t; center_i(t) represents the coordinates of the center point of the i-th vehicle in the three-dimensional world coordinate system; size_i(t) represents the geometric dimensions of the three-dimensional bounding box of the i-th vehicle; heading_i(t) represents the heading angle of the i-th vehicle; v_i(t) represents the instantaneous velocity of the i-th vehicle; a_i(t) represents the instantaneous acceleration of the i-th vehicle; and Δt represents the time interval between two consecutive frames.

3. The method according to claim 1, characterized in that, The step of calculating the road segment-level traffic flow parameters based on the detected motion state parameters of each vehicle includes: The road segment is evenly divided into several sub-segments according to a predetermined length. The traffic flow parameters are calculated for each sub-segment. The formula for calculating the traffic volume Q_seg is Q_seg = N_pass / T_obs. Wherein, N_pass represents the total number of vehicles passing through the cross section of this sub-segment within the preset observation time T_obs; The formula for calculating the average vehicle speed V_avg is V_avg = Σ_{i=1}^{N_v} ||v_i|| / N_v. Where N_v represents the total number of vehicles detected in the current sub-segment, v_i represents the speed vector of the i-th vehicle, and the formula for calculating the lane occupancy rate O_lane is O_lane = Σ_{i=1}^{N_v} l_i / L_seg, where l_i represents the body length of the i-th vehicle, and L_seg represents the segment division length. The queue length L_queue is estimated by detecting the spatial range of continuously moving low-speed vehicles: L_queue = max(s_i) - min(s_i), where ||v_i|| < v_queue, s_i represents the position coordinates of the i-th low-speed vehicle along the road segment direction, and v_queue is the low-speed judgment threshold.

4. The method according to claim 1, characterized in that, The step of calculating micro-traffic behavior indicators based on the detected motion state parameters of each vehicle includes: According to the formula BF = Σ_{i=1}^{N_v} Σ_{t} I(a_i(t) < -a_brake) / (N_v · T_obs), CV_gap = σ_gap / μ_gap, SFI = (1 / N_v) · Σ_{i=1}^{N_v} σ_v_i / μ_v_i is calculated, where I(·) is an indicator function, which takes the value of 1 when the condition in parentheses is met, and takes the value of 0 otherwise. a_i(t) represents the instantaneous acceleration value of the i-th vehicle at time t, a_brake is the braking acceleration judgment threshold, N_v represents the total number of vehicles in the observation range, T_obs represents the observation time length, σ_gap represents the standard deviation of the distance between all adjacent vehicles in the observation segment, μ_gap represents the arithmetic mean of the distance between all adjacent vehicles, σ_v_i represents the standard deviation of the speed of the i-th vehicle in the time dimension, and μ_v_i represents the mean of the speed of the i-th vehicle in the time dimension.

5. The method according to claim 1, characterized in that, The step of combining the traffic flow parameters and the micro-traffic behavior indicators into a traffic state feature vector, and predicting the traffic state for future periods based on the traffic state feature vector, includes: The traffic state feature vector X(t, s) is expressed as: X(t, s) = [Q_seg, V_avg, O_lane, L_queue, BF, CV_gap, SFI]^T, where Q_seg is the traffic flow, V_avg is the average vehicle speed, O_lane is the lane occupancy rate, L_queue is the queue length, BF is the braking frequency, CV_gap is the coefficient of variation of vehicle spacing, and SFI is the speed fluctuation index; The traffic state for future time periods is predicted using the formula X_pred(t+T_p, s) = STTransformer({X(t-T_h, s), ..., X(t, s)}), where X_pred(t+T_p, s) represents the traffic state feature vector at location s on the road segment at time t+T_p in the future; STTransformer(·) represents the spatiotemporal Transformer network model; {X(t-T_h,s), ..., X(t, s)} represents the continuous feature vector sequence within the historical observation window; T_h represents the length of the historical observation window; and T_p represents the prediction time span.

6. A dynamic traffic flow perception and congestion early warning system based on an urban traffic inspection robot, characterized in that, include: The acquisition module is used to acquire multimodal data at a predetermined acquisition frequency. The multimodal data includes three-dimensional point cloud data, image data, positioning data, and the travel speed of the inspection robot itself. The detection module is used to detect vehicles around the inspection robot based on the three-dimensional point cloud data and estimate the motion state parameters of each detected vehicle, including the three-dimensional position, velocity and acceleration of each vehicle. The calculation module is used to calculate the traffic flow parameters at the road segment level based on the motion state parameters of each detected vehicle. The calculation module is also used to calculate micro-traffic behavior indicators based on the detected motion state parameters of each vehicle. The prediction module is used to combine the traffic flow parameters and the micro-traffic behavior indicators into a traffic state feature vector, and predict the traffic state for future periods based on the traffic state feature vector.

7. The system according to claim 6, characterized in that, The detection module is specifically used to calculate based on the formulas V_i(t) = {center_i(t), size_i(t), heading_i(t), v_i(t), a_i(t)}, v_i(t) = (center_i(t) - center_i(t-Δt)) / Δt, and a_i(t) = (v_i(t) - v_i(t-Δt)) / Δt, where V_i(t) represents the state information vector of the i-th vehicle detected at time t; center_i(t) represents the coordinates of the center point of the i-th vehicle in the three-dimensional world coordinate system; size_i(t) represents the geometric dimensions of the three-dimensional bounding box of the i-th vehicle; heading_i(t) represents the heading angle of the i-th vehicle; v_i(t) represents the instantaneous velocity of the i-th vehicle; a_i(t) represents the instantaneous acceleration of the i-th vehicle; and Δt represents the time interval between two consecutive frames.

8. The system according to claim 6, characterized in that, The calculation module is specifically used to uniformly divide the road segment into several sub-segments according to a predetermined length, and calculate the traffic flow parameters for each sub-segment. The traffic flow Q_seg is calculated as Q_seg = N_pass / T_obs, where N_pass represents the total number of vehicles passing through the cross-section of the sub-segment within the preset observation time T_obs; the average vehicle speed V_avg is calculated as V_avg = Σ_{i=1}^{N_v} ||v_i|| / N_v, where N_v represents the total number of vehicles detected in the current sub-segment, and v_i represents the speed vector of the i-th vehicle; the lane occupancy O_lane is calculated as O_lane = Σ_{i=1}^{N_v} l_i / L_seg, where l_i represents the vehicle length of the i-th vehicle, and L_seg represents the segment division length; the queue length L_queue is estimated by detecting the spatial range of continuously slow-moving vehicles: L_queue = max(s_i) - min(s_i), where ||v_i|| < v_queue, s_i represents the position coordinates of the i-th low-speed vehicle along the road segment direction, and v_queue is the low-speed judgment threshold.

9. The system according to claim 6, characterized in that, The calculation module is specifically used to calculate BF = Σ_{i=1}^{N_v} Σ_{t} I(a_i(t) < -a_brake) / (N_v · T_obs), CV_gap = σ_gap / μ_gap, and SFI = (1 / N_v) · Σ_{i=1}^{N_v} σ_v_i / μ_v_i is calculated, where I(·) is an indicator function, which takes the value of 1 when the condition in parentheses is met, and takes the value of 0 otherwise. a_i(t) represents the instantaneous acceleration value of the i-th vehicle at time t, a_brake is the braking acceleration judgment threshold, N_v represents the total number of vehicles in the observation range, T_obs represents the observation time length, σ_gap represents the standard deviation of the distance between all adjacent vehicles in the observation segment, μ_gap represents the arithmetic mean of the distance between all adjacent vehicles, σ_v_i represents the standard deviation of the speed of the i-th vehicle in the time dimension, and μ_v_i represents the mean of the speed of the i-th vehicle in the time dimension.

10. The system according to claim 6, characterized in that, The prediction module is specifically used to represent the traffic state feature vector X(t, s) as: X(t, s) = [Q_seg,V_avg, O_lane, L_queue, BF, CV_gap, SFI]^T, where Q_seg is the traffic flow, V_avg is the average vehicle speed, O_lane is the lane occupancy rate, L_queue is the queue length, BF is the braking frequency, CV_gap is the coefficient of variation of vehicle spacing, and SFI is the speed fluctuation index; according to the formula X_pred(t+T_p, s) = STTransformer({X(t-T_h, s), ...,X(t, s)}), the traffic state in the future time period is predicted, where X_pred(t+T_p, s) represents the predicted traffic state feature vector at the road segment location s at the future time t+T_p; STTransformer(·) represents the spatiotemporal Transformer network model, {X(t-T_h, s), ..., X(t, s)} represents a continuous sequence of feature vectors within a historical observation window; T_h represents the length of the historical observation window; and T_p represents the prediction time span.

Images