Evidence obtaining method and system for right-turn non-stop behavior of national and provincial road truck

Abstract

According to the national and provincial road truck right-turn non-stop behavior evidence obtaining method provided by the invention, the video and picture data collected by the monitoring equipment are utilized to identify the right-turn guide lane boundary and the positioning stop line position of the intersection to be monitored, and then the right-turn guide lane boundary and the positioning stop line position are identified according to the position relationship between the truck and the right-turn guide lane boundary and the positioning stop line position. Judging whether the truck has a line-crossing non-parking behavior or not; three conditions for triggering evidence picture shooting are set, and once the truck continuously triggers three times of shooting, the truck is marked as a suspected vehicle; video stream data which is collected by monitoring equipment and comprises the complete process of the right-turn non-stop behavior of the suspected vehicle, three evidence obtaining photos and a vehicle license plate image form an evidence picture.

Description

Technical Field

[0001] This invention relates to the field of intelligent traffic control technology, specifically to a method and system for collecting evidence of non-stop right turns by large trucks on national and provincial highways. Background Technology

[0002] National and provincial highways are major freight arteries, with a high proportion of heavy truck traffic, reaching over 5,000 vehicles per day on some sections. Due to the large wheel difference and long braking distance of heavy trucks, the failure to stop when turning right has become a major cause of traffic accidents at intersections. According to departmental statistics, over 70% of truck-related injuries and fatalities at traffic light-controlled intersections on national and provincial highways are directly related to trucks failing to stop and observe the movements of non-motorized vehicles and pedestrians when turning right.

[0003] In addressing the issue of large trucks turning right without stopping, early enforcement relied on manual checkpoints. This approach was not only costly and limited by enforcement hours, but also struggled to cover peak violation periods such as early morning, heavy rain, and dense fog. Furthermore, enforcement officers had to observe truck behavior at close range, posing safety risks to themselves. Moreover, during manual enforcement, issues such as staff positioning could lead to subjective judgments of whether a truck had come to a complete stop, resulting in missed or incorrect assessments. For example, a truck briefly coasting might be mistaken for a complete stop, or obstruction of view by non-motorized vehicles could reduce the accuracy rate of violation detection to less than 60%.

[0004] With the widespread use of surveillance equipment, most electronic devices on the market now focus on common violations such as running red lights, speeding, and not driving in designated lanes. However, there is still a lack of solutions specifically designed to identify "large trucks turning right without stopping." This means that even if electronic devices capture large trucks turning right without stopping, the images do not conform to the image rounding specifications stipulated in GA / T832. Furthermore, when multiple large trucks are present, it is impossible to collect targeted evidence data from the violating vehicles. Summary of the Invention

[0005] To address the problem that existing technologies still rely primarily on manual evidence collection for non-stop right-turning behavior of large trucks, this invention provides a method for collecting evidence of such behavior on national and provincial highways. This method enables automated evidence collection, saving manpower and allowing for uninterrupted operation. Furthermore, it can accurately distinguish and collect evidence even when multiple trucks are present. This application also discloses a system for collecting evidence of non-stop right-turning behavior of large trucks on national and provincial highways.

[0006] The technical solution of this invention is as follows: a method for collecting evidence of non-stop right turns by large trucks on national and provincial highways, characterized by comprising the following steps: S1: Install monitoring equipment at the intersection to be monitored; The monitoring equipment includes: electronic police cameras and checkpoint cameras; The electronic police camera collects surveillance video streams, and the checkpoint camera collects images of passing vehicles; S2: Identify traffic elements at the intersection to be monitored; The traffic elements include: the boundary of the right-turn guide lane and the location of the stop line; S3: Identify the trucks based on the video stream data collected by the monitoring equipment, and record the successfully identified trucks as: trucks to be tracked; assign a unique ID to each truck to be tracked, recorded as: tracking ID; S4: Continuously track the vehicle's movement trajectory for each of the aforementioned trucks to be tracked; S5: When a truck to be tracked is detected to be traveling within the boundary of the right-turn guide lane and has not yet reached the positioning stop line, it indicates that the truck has a tendency to turn right. The pre-triggering stage is then initiated and the truck is recorded as the target vehicle. The first frame of evidence image is captured for the target vehicle that enters the pre-triggering phase, recording the state in which the vehicle has not crossed the line; S6: For each target vehicle that completes the first frame capture, mark it as entering the process determination stage; For each target vehicle entering the process determination stage, the trajectory of the truck is continuously tracked. If the following two conditions are met at the same time, the second frame of evidence image is captured to record the target vehicle's behavior of crossing the line without stopping. Condition 1: The vehicle has crossed the stop line; Condition 2: After crossing the stop line, the vehicle remains in motion; S7: For all target vehicles that have completed the second frame capture, continue to track the vehicle's movement trajectory. When the following violation confirmation conditions are met, trigger the capture of the third frame evidence image to confirm the right turn without stopping behavior and mark the target vehicle as: suspect vehicle. The conditions for confirming the violation are: the entire body of the truck has entered the right-hand lane of the motor vehicle; S8: Based on the cross-camera target matching algorithm, the evidence data of the same truck collected by different monitoring devices are matched to synthesize an evidence image for the suspect vehicle. The evidence images include: three frames of evidence images corresponding to the suspect vehicle and a close-up image of the vehicle license plate; S9: Extract the time point when the second frame of evidence image of the suspect vehicle was captured, and record it as: violation trigger time point; Confirm the time period for extracting evidence videos. The extraction time period includes: t1 seconds before the violation is triggered to t2 seconds after the violation is triggered. Based on the tracking ID of the suspected vehicle, video stream data is extracted from the video data of the suspected vehicle captured by the monitoring equipment according to the extraction time period to obtain evidence video. S10: Implement anti-tampering processing for evidence data; Calculate the hash values ​​of the evidence image and the evidence video respectively, and then package and store the hash values ​​with the evidence.

[0007] Its further features are: Step S3 includes the following steps in detail; a1: Construct a coordinate system with the lower left corner of the video frame image captured by the electronic police camera as the origin, and the X-axis parallel to the stop line; a2: Construct a vehicle detection model based on the YOLOv5s model; The vehicle detection model takes video data collected by electronic police cameras as input and outputs the vehicle bounding box, category, confidence score, and coordinates of key points on the axle side of the detected truck in the video. The coordinates of the key points on the vehicle axle side include: point A on the left side of the front axle, point B on the right side of the front axle, point C on the left side of the rear axle, and point D on the right side of the rear axle; a3: Calculate the core parameters based on the coordinates of the key points on the vehicle axle side. The core parameters include: axle center, vehicle center of gravity G, and vehicle bottom bounding box. The axis center includes: front axis center O1 and rear axis center O2; The center O1 of the front axle is the midpoint of the coordinates of points A and B: O1 = ((A x +B X ) / 2, (A y +B y ) / 2); The rear axle center O2 is the midpoint of the coordinates of points C and D: O2 = ((C x +D x ) / 2, (C y +D y ) / 2); The center of gravity G of the vehicle body is the midpoint of the line connecting O1 and O2: G = ((O1, O2) 1x +O 2x ) / 2, (O 1y +O 2y ) / 2); The bottom boundary box of the vehicle body is a rectangular box fitted with four points A, B, C, and D as vertices. The left boundary of the bottom boundary box of the vehicle body is the minimum X value between points A and C, the right boundary is the maximum X value between points B and D, the lower boundary is the minimum Y value among the four points, and the upper boundary is the maximum Y value among the four points. a4: The trucks whose core parameters are successfully detected are denoted as: trucks to be tracked; Using the rear axle center O2 as the core tracking reference, the vehicle trajectory of each of the trucks to be tracked is tracked; a5: Using Vehicle ReID technology, for the video data collected by the electronic police camera, extract the vehicle appearance feature vector for each of the trucks to be tracked, record it as the ReID feature vector, and assign the tracking ID to each truck to be tracked; In step S5, the pre-trigger stage determination method includes the following details: b1: When a target vehicle is detected traveling within the boundary of the right-turn lane, record the tracking ID corresponding to the truck. b2: Obtain the stop line coordinates [P1, P2], where P1 and P2 are the left and right endpoints of the stop line. Since the stop line is parallel to the X-axis, mark the Y-axis coordinates of the stop line as: P y ; b3: Real-time tracking of the truck's trajectory, comparing the target vehicle's coordinates with the stop line's position, and triggering the first frame capture when both of the following non-crossing conditions are met simultaneously, recording the vehicle's non-crossing status; Condition 1 for not crossing the line: Y coordinate of rear axle center O2: O 2y <Stop line y-axis coordinate P> y ; Condition 2 for not crossing the line: The bottom boundary frame of the target vehicle body has no coordinate intersection with the stop line [P1,P2]; In step S6, the trigger determination for capturing the second frame of evidence image specifically includes the following: c1: The method for determining whether the target vehicle has crossed the stop line is as follows: When the Y coordinate of the rear axle center O2 of the target vehicle is from O 2y <P y Completely become O 2y >P y Meanwhile, if the bottom boundary frame of the target vehicle has no coordinate intersection with the stop line [P1,P2], it is determined that the target vehicle has crossed the stop line and entered the right turn area; c2: The method for determining whether the target vehicle is in motion includes the following steps: c21: Record the time when the target vehicle crosses the stop line and enters the right-turn area as: T start ; c22: with T start Starting from the preset Tg seconds of video data, obtain the corresponding video frames to be analyzed; c23: Arrange all video frames to be analyzed in ascending order according to time sequence; obtain the coordinates of the rear axle center O2 and the center of gravity G of the target vehicle in each frame image; c24: Calculate the rear axle center O of the target vehicle in the i-th frame of the video frame to be analyzed. i2x The rear axis center O in the previous frame image i-1 2x Displacement amount: Rear axle center pixel displacement: pd o =O i 2x -O i-1 2x ; c25: Calculate the rear axle center O of the target vehicle in the i-th frame of the video. i 2x Center of gravity of the vehicle body G i x The rear axis center O in the previous frame image i-1 2x Center of gravity of the vehicle body G i-1 x Displacement: Vehicle body center of gravity pixel displacement: pd G =G i x -G i-1 x ; c26: Repeat steps c24~c25 until all video frames to be analyzed have been included in the calculation; c27: When the following two state determination conditions are met simultaneously, the target vehicle is determined to remain in motion. State determination condition 1: All pd o All satisfy: pdo ≥ 2 pixels; State determination condition 2: All pd G All satisfy: pd G ≥2 pixels; In step S7, the method for determining whether the target vehicle body has completely entered the right-hand lane includes the following: d1: Read the x-axis coordinates (Q) of the boundary of the right-turn guide lane. 1x Q 2x ); d2: Track the coordinates of the target vehicle in real time. When the following conditions are met, determine that the truck body has completely entered the right lane and trigger the capture of the third frame of evidence image. The X coordinate O of the rear axle center O2 2x >Q 2x ; In step S8, the cross-camera target matching specifically includes the following steps: c1: For the suspected vehicle, obtain the ReID feature vector of the suspected vehicle extracted by Vehicle ReID, and extract the license plate number and tail feature vector from it. , denoted as: data to be compared; c2 acquires all vehicle passage data collected by the checkpoint camera and extracts the license plate number and vehicle front feature vector. This is recorded as data pending confirmation. c3: Acquire the data to be confirmed one by one, and calculate the cross-camera matching total score S with the data to be compared. total ; c4: Compare S total And the preset similarity threshold Sth; If S total If ≥Sth, then the data to be confirmed and the data to be compared are data from the same vehicle; Otherwise, repeat steps c3 to c4 until all the data to be confirmed has been included in the calculation, then stop the calculation and give a prompt message that no matching data can be found; The cross-mirror matching total score S total The calculation method includes the following steps: c1: Calculate the license plate similarity S plate ; ; In the formula, k represents the k-th character in the license plate; c2: Calculate the similarity S between appearance feature vectors. feat ; Calculate the cosine similarity between the feature vectors of the rear and front of the vehicle: ; The cosine similarity is linearly mapped to a score of 0-30, resulting in S. feat : ; c3: Calculate the spatiotemporal similarity S spatiotemp ; Calculate time similarity S t ; The time it takes for the rear of a vehicle to cross the stop line in the video data collected by the electronic police camera is denoted as: T. 车尾 The vehicle passage time captured by the checkpoint camera is recorded as: T 车头 ; Calculate the time difference ΔT = |T 车头 -T 车尾 |; and compare ΔT with the time threshold Tth; If ΔT≤Tth, then: S t =15×(1-(ΔT / Tth)); Otherwise, if ΔT > Tth, then: S t =0; Calculate spatial similarity S s; Calculate the straight-line distance D between the electronic police camera and the checkpoint camera, and compare D with the preset distance threshold Dth; If D≤Dth, then S s =15×(1-(D / Dth)); Otherwise, if D > Dth, then S s =0; Spatiotemporal similarity S spatiotemp = S t +Ss; c4: Calculate the total cross-camera matching score S total ; Stotal = 0.4 × S plate +0.3×S feat +0.3×S spatiotemp ; The evidence collection images also include two process images captured by a fisheye camera.

[0008] A system for collecting evidence of non-stop right-turning behavior of heavy trucks on national and provincial highways includes: monitoring equipment installed at the intersection to be monitored, the monitoring equipment including: electronic police cameras and checkpoint cameras, characterized in that it further includes: an edge computing device and a network transmission unit, the edge computing device being installed at the intersection to be monitored, and the edge computing device communicating with the electronic police cameras and checkpoint cameras based on the network transmission unit; The edge computing device is equipped with a SOC computing module, connects to the electronic police camera via the RTSP protocol, and connects to the checkpoint camera via the checkpoint SDK; The image acquisition directions of the checkpoint camera and the electronic police camera are respectively facing opposite sides of the roadside sensing area. The checkpoint camera acquires images of the front angle of vehicles at the intersection, and the electronic police camera acquires images of the rear angle of vehicles at the intersection.

[0009] Its further features are: It also includes: a fisheye camera, which is connected to the edge computing device based on the RTSP protocol; The fisheye camera is located between the checkpoint camera and the electronic police camera, with its acquisition direction facing the ground; the image acquisition areas of the fisheye camera, the checkpoint camera, and the electronic police camera cover the entire roadside sensing area.

[0010] This invention provides a method for collecting evidence of trucks making right turns without stopping on national and provincial highways. It utilizes video and image data collected by monitoring equipment to identify the boundary of the right-turn lane and the location of the stop line at the monitored intersection. Based on the positional relationship between the truck and the boundary of the right-turn lane and the stop line, it determines whether the truck has crossed the line without stopping. This application sets three conditions for triggering the capture of evidence images: the first condition is that the truck is traveling within the boundary of the right-turn lane and has not yet reached the stop line; the second condition is that the truck has crossed the line without stopping; and the third condition is that the truck's body has completely entered the right-hand lane. Once the truck triggers three consecutive photographs, it is marked as a suspect vehicle. The video stream data collected by the monitoring equipment, including the complete process of the suspect vehicle's right-turn without stopping, is combined with the three evidence photos and the vehicle license plate image to form the evidence image. The evidence collection process in this application requires no manual intervention and can be implemented continuously. Meanwhile, in this application, a unique tracking ID is assigned to each successfully identified truck to distinguish different vehicles. Based on the target association algorithm, the data corresponding to the same tracking ID is identified, and the vehicle movement trajectory is constructed to ensure that each truck can be accurately tracked and evidence data can be captured even when multiple trucks exist at the same time. Attached Figure Description

[0011] Figure 1 Flowchart of methods for collecting evidence of trucks making right turns without stopping on national and provincial highways; Figure 2 A schematic diagram illustrating the process of a large truck making a right turn without stopping at an intersection to be monitored. Figure 3 A schematic diagram of the system for collecting evidence of non-stop right turns by large trucks on national and provincial highways; Figure 4 This is an example of tracking a target vehicle based on short video frames collected by monitoring equipment. Detailed Implementation

[0012] To reduce system implementation costs, the method for collecting evidence of non-stop right turns by large trucks on national and provincial highways in this application primarily utilizes existing equipment, making the most of devices already installed on the roads. Specifically, as follows... Figure 3 As shown in the figure, the system for collecting evidence of non-stop right-turning behavior of heavy trucks on national and provincial highways constructed in this application includes: monitoring equipment installed on pole 1 at the intersection to be monitored, including: a checkpoint camera 3 and an electronic police camera 2. To ensure real-time calculation, an edge computing device 6 and a network transmission unit (not marked in the figure) are also installed at the intersection. The edge computing device 6 is typically housed in a cabinet at the intersection to be monitored. The edge computing device communicates with the electronic police camera and the checkpoint camera via the network transmission unit. In practice, the checkpoint camera 3, the electronic police camera 2, and the fisheye camera 4 may be installed on the same pole or on different poles. Figure 3 In this embodiment, in order to express the difference in shooting direction of different monitoring devices, the checkpoint camera 3, the electronic police camera 2, and the fisheye camera 4 are all mounted on the same pole.

[0013] Edge computing device 6 is equipped with a SOC computing module. It connects to traffic enforcement camera 2 (1080P / 25fps) via the RTSP protocol (latency ≤500ms) and uses images of passing vehicles (resolution ≥2 megapixels) captured by traffic enforcement camera 3, provided by the traffic enforcement equipment manufacturer. Traffic enforcement camera 2 collects video stream data, and traffic enforcement camera 3 captures images of passing vehicles. The intersection's local area network is configured with gigabit bandwidth to ensure zero packet loss and no lag in video stream and image transmission. Edge computing device 6 includes a perception access layer, an algorithm analysis layer, an evidence generation layer, and a data push layer, deployed within a dedicated video network, and interconnected via a gigabit local area network.

[0014] The core function of the sensing access layer of the edge computing device 6 is to realize data access and image optimization based on technologies such as RTSP / SDK protocol, CLAHE algorithm, and Retinex enhancement.

[0015] The algorithm analysis layer of edge computing device 6 performs the following: Traffic element recognition: The road scene is structured and analyzed. Through edge detection, Hough transform and CNN model, lane dividing lines, stop lines, zebra crossings and traffic light colors and countdown status are automatically recognized. Vehicle detection and tracking: Based on the YOLOv5 model, identify large trucks in videos, extract structured information such as vehicle type, color, and license plate, and use a ReID model trained with Triplet Loss to assign a unique ID to each vehicle and record its movement trajectory; Cross-camera target matching: Multi-dimensional weighted calculation is performed by combining vehicle spatiotemporal information and attribute feature values. When the similarity is ≥90%, it is confirmed that the target captured by the electronic police, checkpoint, and fisheye equipment is the same vehicle.

[0016] The algorithm analysis layer executes a three-stage logic for determining the right-turn non-stop behavior of large trucks: Phase 1: Detect whether there are large trucks within the stop line of the right-turn lane; if so, capture the first process image. Second stage: Determine whether the truck came to a complete stop while turning right after crossing the stop line. If it did not come to a complete stop, take a second picture of the process. The third stage: after the truck crosses the stop line and moves into the right lane, the third process image is captured; if all three stages are met, it is judged as "right turn without stopping".

[0017] The evidence generation layer of edge computing device 6 performs the following operations: Image synthesis: Combine 3 images of the right-turn non-stop behavior process and 1 close-up image (or fisheye process image) of the checkpoint license plate into a "four-in-one" or "six-in-one" evidence image that conforms to GA / T832 or other standards, and overlay the behavior determination time (the time when the second evidence image was taken), location, code, license plate, equipment code and anti-counterfeiting code. Video recording: Record at least 15 seconds of 1080P high-definition video of the right turn without stopping behavior from 5 seconds before the second evidence image is taken to 10 seconds after the trigger, and burn watermark information; Anti-tampering processing: SHA-256 hash values ​​are calculated for image and video files to ensure the integrity of the evidence.

[0018] The data push layer of edge computing device 6 supports two output paths: First, evidence is pushed to the network integration command platform through the upload interface of the centralized command access service for off-site data entry. Second, it connects to the cloud-based alert system to provide real-time warnings to target vehicles via intelligent voice reminders.

[0019] Typically, the image acquisition directions of checkpoint camera 3 and electronic police camera 2 are directed towards opposite sides of the roadside sensing area. Checkpoint camera 3 acquires images of the front angle of vehicles at the intersection, accurately capturing license plate numbers, while electronic police camera 2 acquires images of the rear angle of vehicles at the intersection, accurately determining vehicle behavior. Nighttime video images are preprocessed using the CLAHE algorithm, combined with the Retinex algorithm for feature enhancement. Image correction algorithms are used to optimize substandard image quality, ensuring clear vehicle outlines, license plates, and other key information.

[0020] Normally, if the poles of the checkpoint camera 3 and the traffic enforcement camera 2 are set high enough, the images captured by both can be identified as belonging to the same vehicle based on their exterior features. However, because large trucks are long, if the images taken by the checkpoint camera 3 and the traffic enforcement camera 2 at some intersections are insufficient to identify the same vehicle from the front and rear angles, then a fisheye camera 4 for barefoot scene coverage needs to be installed. The fisheye camera captures a real-time video stream from a panoramic perspective.

[0021] Fisheye camera 4 is connected to edge computing device 6 based on the RTSP protocol; fisheye camera 4 is located between checkpoint camera 3 and electronic police camera 2, with its acquisition direction facing the ground; the image acquisition area of ​​fisheye camera 4, checkpoint camera 3, and electronic police camera 4 covers the entire roadside sensing area. The images acquired by fisheye camera are combined with the images from checkpoint camera 3 and electronic police camera 4 to ensure that a complete evidence image is finally formed.

[0022] Based on image data collected by checkpoint cameras and electronic police cameras, the edge computing device 6 uses YOLOv5 vehicle detection and ReID cross-camera matching technology, along with a three-stage judgment logic for trucks turning right without stopping, to ultimately generate an evidence chain compliant with GA / T832 or other standards. It seamlessly integrates with the command platform and cloud sentry system. A single device supports up to 8 video streams, achieving an accuracy rate of ≥90% in recognizing and matching trucks turning right without stopping on national and provincial highways. This meets the high-precision, 24 / 7 off-site evidence collection requirements for such behavior. This application supports the access of multiple types of equipment, including electronic police cameras, checkpoints, and fisheye cameras, adapting to different national and provincial highway intersection scenarios and reducing modification and deployment costs.

[0023] The evidence collection method based on the above-mentioned method for collecting evidence of trucks making right turns without stopping on national and provincial highways includes the following steps.

[0024] S1: Install monitoring equipment at the intersection to be monitored; The monitoring equipment includes: electronic police cameras and checkpoint cameras; the electronic police cameras capture surveillance video streams, and the checkpoint cameras capture images of passing vehicles. Fisheye cameras may be added as needed to capture real-time video streams with a panoramic view.

[0025] In practice, the CLAHE and Retinex algorithms are also used to optimize the quality of low-light images, ensuring that the recognition accuracy is ≥90% in complex environments such as night and rain, so as to meet the needs of evidence collection around the clock.

[0026] Nighttime video processing: First, the CLAHE algorithm is used to perform adaptive histogram equalization on the image to suppress noise amplification; then, the Retinex algorithm is used to separate the illumination component and reflection component of the image to enhance the details of vehicle outlines, license plates and other features, and improve the signal-to-noise ratio of the image in low light environment by ≥15dB.

[0027] Quality Inspection and Correction: The collected image / video frames are inspected for clarity (sharpness value ≥80) and license plate integrity (unobstructed area ≤10%). Unqualified materials are optimized through image correction algorithms (such as perspective transformation and deblurring) to ensure that they meet the requirements of subsequent analysis.

[0028] S2: Identify traffic elements at the intersection to be monitored; traffic elements include: the boundary of the right-turn guide lane and the location of the stop line.

[0029] There are many existing methods for identifying lane lines and stop lines. In practice, it's sufficient to use existing technologies to identify traffic elements. In this embodiment, lane line and stop line identification uses "edge detection (Canny operator) + Hough transform" to extract the road marking contours. Combined with a CNN semantic segmentation model (such as U-Net), it distinguishes the right-turn guide lane boundary (solid line / dashed line), locates the stop line position, and outputs the coordinates of the left endpoint P1 and right endpoint P2 of the stop line, forming the stop line coordinate range [P]. 1, [P2], error ≤ 5 pixels, such as Figure 2 As shown.

[0030] Simultaneously, a unified video coordinate system is established, with the lower left corner of the intersection as the origin, the horizontal axis as the X-axis, and the vertical axis as the Y-axis. This ensures that the coordinates of the vehicle's four axle-side key points (A / B / C / D) and the derived O1 / O2 / G coordinates can be directly compared and calculated with traffic element coordinates. To simplify calculations, the X-axis is parallel to the stop line in this application. Figure 2 As shown in the figure, the X-axis coordinate of the left boundary line of the right turn guide lane is Q1, and the X-axis coordinate of the right boundary line is Q2.

[0031] S3: Identify large trucks based on video stream data collected by monitoring equipment, and record successfully identified large trucks as: trucks to be tracked; assign a unique ID to each truck to be tracked, recorded as: tracking ID.

[0032] Step S3 includes the following steps in detail; a1: Construct a coordinate system with the upper left corner of the video frame image captured by the electronic police camera as the origin, and the X-axis parallel to the stop line.

[0033] a2: A vehicle detection model is built based on a lightweight improved version of the YOLOv5s model; real-time inference at 25fps is achieved on an edge device to detect large trucks in the scene; the large trucks in this application include: heavy trucks and semi-trailers.

[0034] The vehicle detection model takes video data collected by electronic police cameras as input and outputs the bounding box (8), category, confidence score, and axle-side keypoint coordinates of the detected trucks in the video. The model also outputs structured vehicle information, including license plate number, color, and vehicle model. The axle-side keypoint coordinates include: left front axle point A, right front axle point B, left rear axle point C, and right rear axle point D. The localization error of a single keypoint is ≤3 pixels. Specifically, during the training process of the vehicle detection model, through validation with 1000 sets of labeled samples, the accuracy of simultaneous detection of 4 points is ≥98%.

[0035] a3: Calculate core parameters based on the coordinates of key points on the vehicle axle side. Core parameters include: axle center, vehicle center of gravity G, and vehicle bottom bounding box 11. The axle centers include: front axle center O1 and rear axle center O2; The center O1 of the front axle is the midpoint of the coordinates of points A and B: O1 = ((A x +B X ) / 2, (A y +B y ) / 2); The rear axle center O2 is the midpoint of the coordinates of points C and D: O2 = ((C x +D x ) / 2, (C y +D y ) / 2); The center of gravity G of the vehicle body is the midpoint of the line connecting O1 and O2: G = ((O 1x +O 2x ) / 2, (O 1y +O 2y ) / 2); The bottom bounding box 11 of the vehicle body is a rectangular box fitted with four points A, B, C, and D as vertices. The left boundary of the bottom bounding box 11 of the vehicle body is the minimum X value between points A and C, the right boundary is the maximum X value between points B and D, the lower boundary is the minimum Y value among the four points, and the upper boundary is the maximum Y value among the four points.

[0036] Refer to the instruction manual appendix Figure 4 An example is a video frame from an intersection captured by a traffic enforcement camera. It can be seen that because there is a certain angle between the traffic enforcement camera and the intersection being filmed, from a two-dimensional image perspective of the video frame, the green vehicle bounding boxes 8 corresponding to the two vehicles stopped at the intersection, when projected onto the ground, intersect with the stop line. Therefore, this application does not use the vehicle bounding boxes 8 as a reference for determining the specific position of the vehicles. Instead, it uses the position of the red bottom bounding box 11 of the vehicle body to determine the positional relationship between the vehicle and the traffic elements at the intersection, thereby determining the actual position of the vehicles and ensuring a more accurate final calculation result.

[0037] a4: Trucks whose core parameters have been successfully detected are denoted as: Trucks to be tracked; Using the rear axle center O2 as the core tracking reference, the vehicle trajectory of each truck to be tracked is tracked.

[0038] a5: Using Vehicle ReID technology, for video data collected by electronic police cameras, the vehicle appearance feature vector of each truck to be tracked is extracted and recorded as the ReID feature vector. A tracking ID is then assigned to each truck. The specific Vehicle ReID technology is implemented based on existing technology.

[0039] Please refer to the attached diagram in the instruction manual for details. Figure 4 Each vehicle has a green rectangle, designated as the vehicle bounding box (8), and a red box, designated as the bottom bounding box (11). The axle-side key points (10) are dynamically identified during vehicle operation, so errors may appear in screenshots taken at any given moment. The system uses Vehicle ReID technology to identify and track each vehicle based on its appearance features. Each green vehicle bounding box (8) has a marker (9) at its top left: IDXXXX, which is a unique ID assigned to each vehicle by the system to distinguish it from other vehicles. Figure 4 The green lane line 7 can also be seen. Based on the identification of the lane line, the positional relationship between the vehicle and the lane line can be calculated in real time, thereby judging the vehicle's driving behavior.

[0040] This application utilizes Vehicle Re-Identification (ReID) based on image data acquired by non-overlapping cameras to perform cross-scene identification and tracking of specific vehicles. Vehicle ReID technology identifies vehicles by leveraging their appearance features, such as body texture, window shape, body color, vehicle type, and decorations. Technically, a Vehicle ReID system typically comprises three modules: a feature extraction network employing architectures like ResNet and Vision Transformer to extract multi-scale features; a metric learning module optimizing the feature space distribution using loss functions such as Triplet Loss and Circle Loss; and a retrieval and ranking system that uses cosine similarity or Euclidean distance for rapid candidate set selection, combined with Kalman filtering to predict vehicle trajectories. The specific vehicle ReID system is implemented based on existing technologies. The model is trained using multi-view samples from large trucks to ensure cross-view comparability of front-to-rear features.

[0041] In this application, the rear axle center O2 is used as the core tracking reference because the axle center position is not affected by vehicle body occlusion, making it more suitable for application in scenarios with long wheelbases of large trucks, ensuring more accurate recognition results. This method assigns a unique ID to each large truck to solve the target association problem in occlusion (occlusion rate ≤40%) and cross-view scenarios; simultaneously, it records the coordinate changes of the front axle center O1 and the vehicle's center of gravity G in real time, providing accurate positional basis for subsequent judgments.

[0042] S4: Continuously track the movement trajectory of each truck to be tracked.

[0043] This application designs three stages for recognizing the right-turning non-stop behavior of large trucks: the first stage (pre-triggering), the second stage (process judgment), and the third stage (confirmation of right-turning non-stop behavior), as detailed below.

[0044] S5: The first stage (pre-trigger): When it is detected that the large truck to be tracked is driving inside the boundary of the right-turn guiding lane and has not reached the positioning stop line, indicating that the large truck has the tendency to turn right, it enters the pre-trigger stage, and it is denoted as: the target vehicle, and its tracking ID is locked. Take the first-frame evidence picture of the target vehicle entering the pre-trigger stage, and record the state that the vehicle has not crossed the line.

[0045] Specifically, the determination method in the pre-trigger stage includes the following details: b1: When it is detected that the target vehicle is driving inside the boundary of the right-turn guiding lane, record the tracking ID corresponding to the large truck; b2: Obtain the stop line coordinates [P1, P2], where P1 and P2 are the left and right endpoints of the stop line. Since the stop line is parallel to the X-axis, record the Y-axis coordinate of the stop line as: P y ; b3: Continuously track the trajectory of the large truck, compare the position relationship between the coordinates of the target vehicle and the stop line. When the following two non-crossing conditions are met simultaneously, trigger the first-frame capture and record the state that the vehicle has not crossed the line; Non-crossing condition 1: The Y coordinate of the rear axle center O2: O 2y <The Y-axis coordinate of the stop line P y ; Non-crossing condition 2: There is no coordinate intersection between the bottom boundary box of the target vehicle and the stop line [P1, P2].

[0046] For the target vehicle in the first stage, specifically refer to Figure 2 In, the position of the large truck marked as 1, the vehicle is between the left and right lane lines of the left-turn guiding lane boundary line, that is: Q1 < rear axle center O2 < Q2; because the coordinate origin in this application is set at the lower left corner, the larger the y value of the point farther from the origin, and the vehicle is below the stop line, so O 2y <The Y-axis coordinate of the stop line P y , and at the same time, the bottom boundary box of the vehicle has no intersection with the stop line completely.

[0047] S6: The second stage (process determination): For each target vehicle that has completed the first-frame capture, it is marked as entering the process determination stage; For each target vehicle entering the process determination stage, continuously track the trajectory of the large truck. If the following two conditions are met simultaneously, trigger the second-frame evidence picture capture and record the behavior of the target vehicle crossing the line without stopping; Condition 1: The vehicle has crossed the stop line; Condition 2: After crossing the stop line, the vehicle remains in a moving state.

[0048] In step S6, the trigger determination for capturing the second frame of evidence image specifically includes the following: c1: The method for determining whether the target vehicle has crossed the stop line is as follows: When the Y-coordinate of the rear axle center O2 of the target vehicle changes from O 2y <P y Completely become O 2y >P y Meanwhile, if the bottom bounding box of the target vehicle has no coordinate intersection with the stop line [P1,P2], it is determined that the target vehicle has crossed the stop line and entered the right turn area. c2: The method for determining whether the target vehicle is in motion includes the following steps: c21: Record the time when the target vehicle crosses the stop line and enters the right-turn area as: T start ; c22: with T start Starting from the preset Tg seconds of video data, obtain the corresponding video frames to be analyzed; c23: Sort all video frames to be analyzed in ascending order of time; obtain the coordinates of the rear axle center O2 and the center of gravity G of the target vehicle in each frame; c24: Calculate the rear axle center O of the target vehicle in the i-th frame of the video frame to be analyzed. i 2x The rear axis center O in the previous frame image i-1 2x Displacement amount: Rear axle center pixel displacement: pd o =O i 2x -O i-1 2x ; c25: Calculate the rear axle center O of the target vehicle in the i-th frame of the video. i 2x Center of gravity of the vehicle body G i x The rear axis center O in the previous frame image i-1 2x Center of gravity of the vehicle body G i-1 x Displacement: Vehicle body center of gravity pixel displacement: pd G =G i x -G i-1 x ; c26: Repeat steps c24~c25 until all video frames to be analyzed have been included in the calculation; c27: When the following two state determination conditions are met simultaneously, the target vehicle is determined to remain in motion. State determination condition 1: All pd o All satisfy: pdo ≥ 2 pixels; State determination condition 2: All pd G All satisfy: pd G ≥2 pixels; The specific Tg value is set according to the frame rate of the electronic police camera. In this embodiment, the Tg value is set to 0.4 seconds, and 0.4 seconds of monitoring video data corresponds to 10 video frames.

[0049] The target vehicles in the second phase refer to the appendix. Figure 2 The truck marked 2 is above the stop line compared to the truck marked 1, but the target vehicle is still on the left side of the right extension of the lane.

[0050] S7: Third stage (confirmation of right turn without stopping behavior): For all target vehicles that have completed the second frame capture, continue to track the vehicle's movement trajectory. When the following violation confirmation conditions are met, trigger the capture of the third frame evidence image to confirm the right turn without stopping behavior and mark the target vehicle as: suspect vehicle.

[0051] The conditions for confirming a violation are: the entire body of the large truck has entered the right-hand lane of the motor vehicle.

[0052] In step S7, the method for determining whether the target vehicle body has completely entered the right-hand lane includes the following: d1: Read the x-axis coordinates (Q) of the boundary of the right-turn guide lane. 1x Q 2x ); d2: Track the coordinates of the target vehicle in real time. When the following conditions are met, determine that the truck body has completely entered the right lane and trigger the capture of the third frame of evidence image. The X coordinate O of the rear axle center O2 2x >Q 2x .

[0053] Refer to the attached diagram in the instruction manual. Figure 2 The truck marked 3 is positioned entirely to the right of the right-hand extension of the lane. In this embodiment, the vehicle's position is represented by its rear axle center O2. 2x >Q 2x This indicates that the target vehicle has completely entered the right-hand lane.

[0054] S8: Based on the cross-camera target matching algorithm, the evidence data of the same truck collected by different monitoring devices are matched to synthesize evidence images of the suspect vehicle.

[0055] The evidence includes three frames of evidence images corresponding to the suspect vehicle and a close-up image of the vehicle's license plate.

[0056] If the system is configured to use a fisheye camera to capture panoramic images, the evidence images will also include two process images captured by the fisheye camera.

[0057] In step S8, cross-camera target matching specifically includes the following steps: c1: Based on the video data collected by the electronic police camera corresponding to the suspect vehicle, obtain the suspect vehicle ReID feature vector extracted by VehicleReID, and extract the license plate number and tail feature vector from it. , denoted as: data to be compared; c2 acquires all vehicle passage data collected by the checkpoint camera and extracts the license plate number and vehicle front feature vector. This is recorded as data pending confirmation. c3: Acquire the data to be confirmed one by one, and calculate the cross-camera matching total score S with the data to be compared. total ; c4: Compare S total And the preset similarity threshold Sth; If S total If ≥Sth, then the data to be confirmed and the data to be compared are data from the same vehicle. Otherwise, repeat steps c3 to c4 until all the data to be confirmed has been included in the calculation, then stop the calculation and give a message that no matching data can be found.

[0058] The similarity threshold Sth is set according to the actual calculation accuracy. In this embodiment, Sth is set to 90 points.

[0059] Cross-camera matching total score S total The calculation method includes the following steps: c1: Calculate the license plate similarity S plate ; Large truck license plates have a 7-digit structure, specifically: province abbreviation + letter + 5-digit number / letter combination.

[0060] When comparing the rear license plate of the electronic police vehicle with the front license plate of the vehicle at the checkpoint, each digit needs to be matched and scored: If each digit is perfectly matched, the score is approximately 40 / 7 ≈ 5.71. If the characters are blurry or partially obscured, each confirmed match scores 40 / 7 × 0.8 = 4.57 points, with the specific score determined by a pre-designed and trained character recognition model.

[0061] The corresponding formula is: ; In the formula, k represents the k-th character in the license plate.

[0062] c2: Calculate the similarity S between appearance feature vectors. feat ; Calculate the cosine similarity between the feature vectors of the rear and front of the vehicle: ; The cosine similarity is linearly mapped to a score of 0-30, resulting in S. feat : .

[0063] c3: Calculate the spatiotemporal similarity S spatiotemp ; Calculate time similarity S t ; The time it takes for the rear of a vehicle to cross the stop line in the video data collected by the electronic police camera is denoted as: T. 车尾 The time it takes for the front of the vehicle to cross the stop line, captured by the checkpoint camera, is recorded as: T 车头 ; Calculate the time difference ΔT = |T 车头 -T 车尾 |; and compare ΔT with the time threshold Tth; the specific value of Tth is set according to the actual situation of the intersection. In this embodiment, Tth is set to 30 seconds; If ΔT≤Tth, then: S t =15×(1-(ΔT / Tth)); Otherwise, if ΔT > Tth, then: S t =0; Calculate spatial similarity S s Based on the pole installation position, calculate the spatial straight-line distance D (unit: meters) between the electronic police camera and the checkpoint camera, and compare D with the preset distance threshold Dth; in this embodiment, Dth is set to 50 meters. If D≤Dth, then S s =15×(1-(D / Dth)); Otherwise, if D > Dth, then S s =0; Spatiotemporal similarity S spatiotemp = S t +Ss.

[0064] c4: Calculate the total cross-camera matching score S total ; S total =0.4×S plate +0.3×S feat +0.3×S spatiotemp .

[0065] In this application, the license plate similarity score is 40 points, with a weight of 40%; the appearance feature vector similarity score is 30 points, with a weight of 30%; and the spatiotemporal similarity score is 30 points, with a weight of 30%. In this application, the ReID model and multi-dimensional weighting algorithm are used to achieve cross-camera target association between electronic police, checkpoint, and fisheye devices, with a matching accuracy of ≥90%, ensuring that the target vehicle is uniquely traceable.

[0066] S9: The time point at which the second frame of evidence image of the suspect vehicle was captured is recorded as: the violation trigger time point; Confirm the time period for extracting evidence videos. The extraction time period includes: t1 seconds before the violation is triggered to t2 seconds after the violation is triggered. Based on the tracking ID of the suspect vehicle, video stream data is extracted from the video data of the suspect vehicle captured by the monitoring equipment according to the extraction time period to obtain evidence video. Finally, violation information is added as a watermark to the synthesized evidence image.

[0067] S10: Implement anti-tampering processing for evidence data; Calculate the hash values ​​of the evidence images and evidence videos separately, and then package and store the hash values ​​along with the evidence.

[0068] Evidence is automatically pushed to the integrated command platform for punishment, or real-time alerts are provided through the cloud sentry system, without the need for manual intervention, forming a closed loop of "evidence collection-disposal".

[0069] The images of vehicles passing through checkpoints serve as "close-up images of the vehicle's license plate," focusing on the license plate information; fisheye video frames serve as "process images for filling blind spots," supplementing details of vehicle movement in blind spots of electronic police / checkpoints (such as vehicle posture and wheel track direction), and assisting in reconstructing the entire process of right-turning without stopping; both together supplement the evidence chain, achieving an evidence loop of "close-up + process + blind spot filling."

[0070] When synthesizing evidence, in accordance with the GA / T832 standard, three process images of evidence taken in three stages and one close-up image of the license plate taken by the checkpoint camera are combined into a "four-in-one" evidence image, or two fisheye process images are added to combine into a "six-in-one" evidence image. The following information is superimposed below the image: time (accurate to milliseconds, such as 2025-10-17 15:30:45.123), location (such as S342 provincial highway K120+500 meters), behavior type code (such as 1116B), specific behavior description (at an intersection with a stop-and-yield sign, a large vehicle turns right without stopping to yield), equipment code (such as 320125000000010001), and anti-counterfeiting code.

[0071] • Video recording: Employs circular buffer technology to save video from 5 seconds before the right turn without stopping is triggered to 10 seconds after the trigger (total duration ≥ 15 seconds), encoded in H.264 format with a resolution of 1080P; depending on actual needs, character information can also be burned into an unremovable watermark using FFmpeg.

[0072] • Anti-tampering processing: Calculate the SHA-256 hash value of the "four-in-one" image and the video of the right turn without stopping, store it in the local database of the edge device, and write it into the evidence package at the same time to ensure that the data has not been tampered with.

[0073] During data push and integration, the system connects to the integrated command platform: Evidence packages (composite images + videos + behavioral descriptions) are pushed to the integrated command platform for off-site data entry via the platform's upload interface (HTTP protocol). It also connects to the cloud-based sentry system: The suspected vehicle's license plate number and information about its right-turn failure are pushed in real-time to the cloud-based sentry intelligent voice terminal via the platform's upload interface (HTTP protocol), triggering a voice reminder of "large truck turning right without stopping," providing immediate alerts.

Claims

1. A method for collecting evidence of non-stop right turns by large trucks on national and provincial highways, characterized in that, It includes the following steps: S1: Install monitoring equipment at the intersection to be monitored; The monitoring equipment includes: electronic police cameras and checkpoint cameras; The electronic police camera collects surveillance video streams, and the checkpoint camera collects images of passing vehicles; S2: Identify traffic elements at the intersection to be monitored; The traffic elements include: the boundary of the right-turn guide lane and the location of the stop line; S3: Identify the trucks based on the video stream data collected by the monitoring equipment, and record the successfully identified trucks as: trucks to be tracked; assign a unique ID to each truck to be tracked, recorded as: tracking ID; S4: Continuously track the vehicle's movement trajectory for each of the aforementioned trucks to be tracked; S5: When a truck to be tracked is detected to be traveling within the boundary of the right-turn guide lane and has not yet reached the positioning stop line, it indicates that the truck has a tendency to turn right. The pre-triggering stage is then initiated and the truck is recorded as the target vehicle. The first frame of evidence image is captured for the target vehicle that enters the pre-triggering phase, recording the state in which the vehicle has not crossed the line; S6: For each target vehicle that completes the first frame capture, mark it as entering the process determination stage; For each target vehicle entering the process determination stage, the trajectory of the truck is continuously tracked. If the following two conditions are met at the same time, the second frame of evidence image is captured to record the target vehicle's behavior of crossing the line without stopping. Condition 1: The vehicle has crossed the stop line; Condition 2: After crossing the stop line, the vehicle remains in motion; S7: For all target vehicles that have completed the second frame capture, continue to track the vehicle's movement trajectory. When the following violation confirmation conditions are met, trigger the capture of the third frame evidence image to confirm the right turn without stopping behavior and mark the target vehicle as: suspect vehicle. The violation confirmation condition is: the entire body of the large truck has entered the right-hand lane of the motor vehicle; S8: Based on the cross-camera target matching algorithm, the evidence data of the same truck collected by different monitoring devices are matched, and an evidence collection image is synthesized for the suspect vehicle. The evidence images include: three frames of evidence images corresponding to the suspect vehicle and a close-up image of the vehicle license plate; S9: Extract the time point when the second frame of the evidence image of the suspect vehicle was captured, and record it as: violation trigger time point; Confirm the time period for extracting evidence videos. The extraction time period includes: t1 seconds before the violation is triggered to t2 seconds after the violation is triggered. Based on the tracking ID of the suspected vehicle, video stream data is extracted from the video data of the suspected vehicle captured by the monitoring equipment according to the extraction time period to obtain evidence video. S10: Implement anti-tampering processing for evidence data; Calculate the hash values ​​of the evidence image and the evidence video respectively, and then package and store the hash values ​​with the evidence.

2. The method for collecting evidence of non-stop right turns by large trucks on national and provincial highways according to claim 1, characterized in that: Step S3 includes the following steps in detail; a1: Construct a coordinate system with the lower left corner of the video frame image captured by the electronic police camera as the origin, and the X-axis parallel to the stop line; a2: Construct a vehicle detection model based on the YOLOv5s model; The vehicle detection model takes video data collected by electronic police cameras as input and outputs the vehicle bounding box, category, confidence score, and coordinates of key points on the axle side of the detected truck in the video. The coordinates of the key points on the vehicle axle side include: point A on the left side of the front axle, point B on the right side of the front axle, point C on the left side of the rear axle, and point D on the right side of the rear axle; a3: Calculate the core parameters based on the coordinates of the key points on the vehicle axle side. The core parameters include: axle center, vehicle center of gravity G, and vehicle bottom bounding box. The axis center includes: front axis center O1 and rear axis center O2; The center O1 of the front axle is the midpoint of the coordinates of points A and B: O1 = ((A x +B X ) / 2, (A y +B y ) / 2); The rear axle center O2 is the midpoint of the coordinates of points C and D: O2 = ((C x +D x ) / 2, (C y +D y ) / 2); The center of gravity G of the vehicle body is the midpoint of the line connecting O1 and O2: G = ((O1, O2) 1x +O 2x ) / 2, (O 1y +O 2y ) / 2); The bottom boundary box of the vehicle body is a rectangular box fitted with four points A, B, C, and D as vertices. The left boundary of the bottom boundary box of the vehicle body is the minimum X value between points A and C, the right boundary is the maximum X value between points B and D, the lower boundary is the minimum Y value among the four points, and the upper boundary is the maximum Y value among the four points. a4: The trucks whose core parameters are successfully detected are denoted as: trucks to be tracked; Using the rear axle center O2 as the core tracking reference, the vehicle trajectory of each of the trucks to be tracked is tracked; a5: Using Vehicle ReID technology, for the video data collected by the electronic police camera, extract the vehicle appearance feature vector for each of the trucks to be tracked, record it as the ReID feature vector, and assign the tracking ID to each truck.

3. The method for collecting evidence of non-stop right turns by large trucks on national and provincial highways according to claim 2, characterized in that: In step S5, the pre-trigger stage determination method includes the following details: b1: When a target vehicle is detected traveling within the boundary of the right-turn lane, record the tracking ID corresponding to the truck. b2: Obtain the coordinates of the stop line [P1, P2], where P1 and P2 are the left and right endpoints of the stop line. Since the stop line is parallel to the X-axis, mark the Y-axis coordinates of the stop line as: P y ; b3: Real-time tracking of the truck's trajectory, comparing the target vehicle's coordinates with the stop line's position, and triggering the first frame capture when both of the following non-crossing conditions are met simultaneously, recording the vehicle's non-crossing status; Condition 1 for not crossing the line: Y coordinate of rear axle center O2: O 2y <Stop line y-axis coordinate P> y ; Condition 2 for not crossing the line: The bottom boundary frame of the target vehicle body has no coordinate intersection with the stop line [P1,P2].

4. The method for collecting evidence of non-stop right turns by large trucks on national and provincial highways according to claim 2, characterized in that: In step S6, the trigger determination for capturing the second frame of evidence image specifically includes the following: c1: The method for determining whether the target vehicle has crossed the stop line is as follows: When the Y coordinate of the rear axle center O2 of the target vehicle is from O 2y <P y Completely become O 2y > P y Meanwhile, if the bottom boundary frame of the target vehicle has no coordinate intersection with the stop line [P1,P2], it is determined that the target vehicle has crossed the stop line and entered the right turn area; c2: The method for determining whether the target vehicle is in motion includes the following steps: c21: Record the time when the target vehicle crosses the stop line and enters the right-turn area as: T start ; c22: with T start Starting from the preset Tg seconds of video data, obtain the corresponding video frames to be analyzed; c23: Arrange all video frames to be analyzed in ascending order according to time sequence; obtain the coordinates of the rear axle center O2 and the center of gravity G of the target vehicle in each frame image; c24: Calculate the rear axle center O of the target vehicle in the i-th frame of the video frame to be analyzed. i 2x The rear axis center O in the previous frame image i-1 2x Displacement amount: Rear axle center pixel displacement: pd o =O i 2x -O i-1 2x ; c25: Calculate the rear axle center O of the target vehicle in the i-th frame of the video. i 2x Center of gravity of the vehicle body G i x The rear axis center O in the previous frame image i-1 2x Center of gravity of the vehicle body G i-1 x Displacement: Vehicle body center of gravity pixel displacement: pd G =G i x -G i-1 x ; c26: Repeat steps c24~c25 until all video frames to be analyzed have been included in the calculation; c27: When the following two state determination conditions are met simultaneously, the target vehicle is determined to remain in motion. State determination condition 1: All pd o All satisfy: pdo ≥ 2 pixels; State determination condition 2: All pd G All satisfy: pd G ≥2 pixels.

5. The method for collecting evidence of non-stop right turns by large trucks on national and provincial highways according to claim 2, characterized in that: In step S7, the method for determining whether the target vehicle body has completely entered the right-hand lane includes the following: d1: Read the x-axis coordinates (Q) of the boundary of the right-turn guide lane. 1x Q 2x ); d2: Track the coordinates of the target vehicle in real time. When the following conditions are met, determine that the truck body has completely entered the right lane and trigger the capture of the third frame of evidence image. The X coordinate O of the rear axle center O2 2x >Q 2x .

6. The method for collecting evidence of non-stop right turns by large trucks on national and provincial highways according to claim 1, characterized in that: In step S8, the cross-camera target matching specifically includes the following steps: c1: For the suspected vehicle, obtain the ReID feature vector of the suspected vehicle extracted by Vehicle ReID, and extract the license plate number and tail feature vector from it. , denoted as: data to be compared; c2 acquires all vehicle passage data collected by the checkpoint camera and extracts the license plate number and vehicle front feature vector. Record this as data pending confirmation; c3: Acquire the data to be confirmed one by one, and calculate the cross-camera matching total score S with the data to be compared. total ; c4: Compare S total And the preset similarity threshold Sth; If S total If ≥Sth, then the data to be confirmed and the data to be compared are data from the same vehicle; Otherwise, repeat steps c3 to c4 until all the data to be confirmed has been included in the calculation, then stop the calculation and give a prompt message that no matching data can be found.

7. The method for collecting evidence of non-stop right turns by large trucks on national and provincial highways according to claim 6, characterized in that: The cross-mirror matching total score S total The calculation method includes the following steps: c1: Calculate the license plate similarity S plate ; ； In the formula, k represents the k-th character in the license plate; c2: Calculate the similarity S between appearance feature vectors. feat ; Calculate the cosine similarity between the feature vectors of the rear and front of the vehicle: ； The cosine similarity is linearly mapped to a score of 0-30, resulting in S. feat : ； c3: Calculate the spatiotemporal similarity S spatiotemp ; Calculate time similarity S t ; The time it takes for the rear of a vehicle to cross the stop line in the video data collected by the electronic police camera is denoted as: T. 车尾 The vehicle passage time captured by the checkpoint camera is recorded as: T 车头 ; Calculate the time difference ΔT = |T 车头 -T 车尾 |; and compare ΔT with the time threshold Tth; If ΔT≤Tth, then: S t =15×(1-(ΔT / Tth)); Otherwise, if ΔT > Tth, then: S t =0; Calculate spatial similarity S s ; Calculate the spatial straight-line distance D between the electronic police camera and the checkpoint camera, and compare D with the preset distance threshold Dth; If D≤Dth, then S s =15×(1-(D / Dth)); Otherwise, if D > Dth, then S s =0; Spatiotemporal similarity S spatiotemp = S t +Ss; c4: Calculate the total cross-camera matching score S total ; Stotal=0.4×S plate +0.3×S feat +0.3×S spatiotemp 。 8. The method for collecting evidence of non-stop right turns by large trucks on national and provincial highways according to claim 1, characterized in that: The evidence collection images also include: two process images captured by a fisheye camera.

9. A system for collecting evidence of non-stop right-turning behavior of large trucks on national and provincial highways, comprising: A monitoring device installed at an intersection to be monitored, the monitoring device including: an electronic police camera and a checkpoint camera, characterized in that it further includes: an edge computing device and a network transmission unit, the edge computing device being installed at the intersection to be monitored, the edge computing device communicating with the electronic police camera and the checkpoint camera based on the network transmission unit; The edge computing device is equipped with a SOC computing module, connects to the electronic police camera via the RTSP protocol, and connects to the checkpoint camera via the checkpoint SDK; The image acquisition directions of the checkpoint camera and the electronic police camera are respectively facing opposite sides of the roadside sensing area. The checkpoint camera acquires images of the front angle of vehicles at the intersection, and the electronic police camera acquires images of the rear angle of vehicles at the intersection.

10. The system for collecting evidence of non-stop right-turning behavior of large trucks on national and provincial highways according to claim 9, characterized in that, It also includes: a fisheye camera, which is connected to the edge computing device based on the RTSP protocol; The fisheye camera is located between the checkpoint camera and the electronic police camera, with its acquisition direction facing the ground; the image acquisition areas of the fisheye camera, the checkpoint camera, and the electronic police camera cover the entire roadside sensing area.

Images