AI-based visual inspection video surveillance system

Abstract

This invention relates to the field of AI vision technology, specifically disclosing a video surveillance system based on AI vision detection, comprising: an environment setting module: acquiring panoramic images to establish a two-dimensional coordinate system, obtaining vehicle speed limits, and calculating the sampling frequency; an image processing module: capturing images at the sampling frequency, acquiring wheel feature points in the vehicle body, recording the coordinates of wheel feature points in each frame of the image, and calculating the coordinates of occluded wheel feature points; and an analysis module: recording the coordinates of wheel feature points in each frame of the image throughout the vehicle's passage through the intersection, generating wheel trajectory curves, and comparing the wheel trajectory curves with prohibitory traffic markings to analyze whether any violations exist. This solution, by establishing a two-dimensional coordinate system, dynamically adjusting the sampling frequency, and combining interpolation completion technology to generate wheel trajectories, distinguishes between crossing and overtaking violations, overcomes environmental limitations, achieves stable all-weather monitoring, improves judgment accuracy and law enforcement credibility, and optimizes traffic control.

Description

Technical Field

[0001] This invention relates to the field of AI vision technology, and more specifically to a video surveillance system based on AI vision detection. Background Technology

[0002] Traffic violations captured by cameras at intersections refer to violations committed by vehicles that cross or over prohibited traffic lines, which are detected by electronic monitoring equipment at intersections. This aims to regulate traffic order and reduce traffic accidents.

[0003] Today, intersection traffic enforcement cameras have been upgraded to high-definition camera systems with extremely high accuracy. Their determination typically requires three key photos: a snapshot of the vehicle's front wheels crossing the solid line, a photo of the rear wheels crossing the solid line for license plate recognition, and a panoramic photo of the vehicle passing through the area crossing the solid line at the intersection. These three photos corroborate each other to confirm the violation.

[0004] However, inclement weather and insufficient light lead to inaccurate recognition. In particular, rainwater reflections can blur road markings, significantly reducing the success rate of image capture. At night, without streetlights or in dim lighting, the contrast between license plates and road markings is low, resulting in an extremely low success rate. In backlit scenes at dawn and dusk, traditional equipment is also prone to overexposure or excessive shadows, making it impossible to determine whether a vehicle is crossing the line. These inclement weather and lighting issues can cause a large number of missed or misjudged images. Summary of the Invention

[0005] The purpose of this invention is to provide a video surveillance system based on AI visual inspection to solve the above-mentioned technical problems.

[0006] The objective of this invention can be achieved through the following technical solutions: AI-based visual inspection video surveillance systems include: Environment Setup Module: Acquire panoramic images of the intersection, establish a two-dimensional coordinate system, where the center point of the intersection is the origin, the horizontal axis represents the transverse direction of the road, and the vertical axis represents the driving direction. Obtain the vehicle speed limit v at the intersection. The method for calculating the sampling frequency F based on the vehicle speed limit v includes: Preset velocity gradient , Let λ represent the preset unit speed, λ represent the preset gradient coefficient, λ=1, 2..., obtain the value of λ when the constraint condition v / V≤1 is met, and let the sampling frequency F=λ×f, where f represents the initial sampling frequency of the intersection camera; The vehicle body in the image is obtained, and the wheel feature points in the vehicle body are obtained. The wheel feature points refer to the points where the wheel contacts the road surface. The coordinates (x, y) of the wheel feature points in each frame are recorded. If the wheel feature point is occluded in a single frame image, obtain the coordinates corresponding to the Nth frame before and the Nth frame after. and Let the coordinates of the obscured wheel feature point be... Where N represents the preset offset frame number, N > 0; Analysis module: Records the coordinates of wheel feature points in each frame of the image during the entire process of a vehicle passing through an intersection in chronological order, generates a wheel trajectory curve L, and compares the wheel trajectory curve L with the prohibitory traffic markings. If the wheel track curve L coincides with the prohibitory traffic marking, then the vehicle has crossed the line when passing through the intersection. If the wheel trajectory curve L crosses the prohibitory traffic markings and the maximum lateral distance exceeds a preset distance threshold, then the vehicle has crossed the line when passing through the intersection.

[0007] As a further aspect of the present invention: in the environment setting module, the traffic flow C at the intersection per unit time is obtained; if the traffic flow C is lower than a preset traffic flow threshold C... min Set the sampling frequency of the intersection camera to 50% of the initial sampling frequency.

[0008] As a further aspect of the present invention: in the image processing module, if the wheel feature points in the Nth frame before and the Nth frame after are occluded, the wheel feature points in the (N+1)th frame before and the (N+1)th frame after are obtained, and it is determined whether the wheel feature points are occluded. If they are not occluded, the process stops; if they are occluded, the above operation is repeated until the wheel feature points are no longer occluded.

[0009] As a further aspect of the present invention: in the analysis module, four wheel trajectory curves corresponding to the vehicle are obtained. When none of the four wheel trajectory curves show any line-crossing or line-exceeding behavior, the vehicle is judged to be driving normally.

[0010] As a further aspect of the present invention: In the analysis module, when construction barriers, temporary markings, water-covered markings, or snow-buried markings appear at an intersection, the temporary markings are identified from the original prohibitory traffic markings, with the temporary markings serving as the criterion for judgment.

[0011] As a further aspect of the present invention: In the analysis module, when the wheel feature points in the image are misaligned, the visible light and infrared light cameras are time-synchronized and calibrated so that the error between the calibrated visible light and infrared light cameras is less than or equal to 1 / F.

[0012] As a further aspect of the present invention: In the analysis module, if the vehicle length exceeds a preset judgment length, it is classified as a large truck, and the following operations are performed on large trucks: Obtain wheel feature points A1, A2...An on the same side, where n represents the number of wheels on one side. Set a virtual feature point a, and the horizontal and vertical coordinates of the virtual feature point a are the average of the corresponding horizontal and vertical coordinates of wheel feature points A1, A2...An. Generate the wheel trajectory curve LA of the virtual feature point a, and compare the wheel trajectory curve LA with the prohibitory traffic markings.

[0013] The beneficial effects of this invention are as follows: This solution establishes a two-dimensional coordinate system at the intersection and dynamically adjusts the sampling frequency F according to the vehicle speed limit to achieve high-density continuous capture of wheel trajectories and completely reconstruct the driving path. In scenarios where wheel feature points are occluded, a coordinate interpolation method of N frames before and after is used to complete the trajectory, avoiding trajectory breaks. By comprehensively comparing the trajectory curve with prohibitory markings, the invention accurately distinguishes between driving over and crossing the line, significantly improving the accuracy and credibility of the judgment.

[0014] This solution overcomes the environmental limitations of traditional equipment, focusing on trajectory feature extraction and comparison rather than solely on image clarity. Dynamic sampling frequency adapts to different vehicle speeds, and coordinate interpolation algorithms compensate for occlusion and image blurring, reducing reliance on lighting and weather conditions. In scenarios such as rainy days with glare, dim nighttime conditions, and backlighting in the morning and evening, only some wheel feature points need to be captured for judgment, achieving stable monitoring around the clock. Simultaneously, the solution quantifies position through a coordinate system, establishing an objective and unified judgment standard, avoiding the subjectivity of human judgment, ensuring consistency in judgment across different vehicle speeds, providing a clear and traceable chain of evidence, and enhancing impartiality. Furthermore, it requires no additional sensors, achieving high-precision judgment solely through cameras and algorithm optimization. Hardware costs are controllable, and upgrades are easy, avoiding data redundancy, improving processing efficiency, and providing detailed violation statistics for traffic management, helping to optimize control strategies and aligning with the core objectives of regulating traffic order and reducing accidents. Attached Figure Description

[0015] The invention will now be further described with reference to the accompanying drawings.

[0016] Figure 1 This is a schematic diagram of the structure of the video surveillance system based on AI visual inspection according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 As shown, the present invention is a video surveillance system based on AI visual inspection, comprising: Environment Setup Module: Acquire panoramic images of the intersection, establish a two-dimensional coordinate system, where the center point of the intersection is the origin, the horizontal axis represents the transverse direction of the road, and the vertical axis represents the driving direction. Obtain the vehicle speed limit v at the intersection. The method for calculating the sampling frequency F based on the vehicle speed limit v includes: Preset velocity gradient , Let λ represent the preset unit speed, λ represent the preset gradient coefficient, λ=1, 2..., obtain the value of λ when the constraint condition v / V≤1 is met, and let the sampling frequency F=λ×f, where f represents the initial sampling frequency of the intersection camera; Image processing module: When a vehicle enters the intersection monitoring range, the intersection camera is instructed to take a picture at a sampling frequency F, and the captured image is processed as follows: The vehicle body in the image is obtained, and the wheel feature points in the vehicle body are obtained. The wheel feature points refer to the points where the wheel contacts the road surface. The coordinates (x, y) of the wheel feature points in each frame are recorded. If the wheel feature point is occluded in a single frame image, obtain the coordinates corresponding to the Nth frame before and the Nth frame after. and Let the coordinates of the obscured wheel feature point be... Where N represents the preset offset frame number, N > 0; Analysis module: Records the coordinates of wheel feature points in each frame of the image during the entire process of a vehicle passing through an intersection in chronological order, generates a wheel trajectory curve L, and compares the wheel trajectory curve L with the prohibitory traffic markings. If the wheel track curve L coincides with the prohibitory traffic marking, then the vehicle has crossed the line when passing through the intersection. If the wheel trajectory curve L crosses the prohibitory traffic markings and the maximum lateral distance exceeds a preset distance threshold, then the vehicle has crossed the line when passing through the intersection.

[0019] It should be noted that, in addition to its core functions, the environment settings module needs to complete two additional key tasks: First, the panoramic image acquisition must cover the entire intersection area, including all prohibitory traffic markings, lane boundaries, intersection edges, and surrounding obstructions; second, after establishing the two-dimensional coordinate system, the prohibitory traffic markings need to be calibrated, converting their physical positions into coordinate intervals within the coordinate system to form a standardized comparison benchmark library. Simultaneously, parameters such as the number of lanes, the width of each lane, and the thickness of the markings must be recorded to provide a reference for the distance threshold used in subsequent lane crossing determination. The calculation of the sampling frequency F must follow the principle of matching speed and accuracy, ensuring that both low-speed and high-speed vehicles are at the most suitable sampling interval, guaranteeing trajectory continuity while avoiding data redundancy.

[0020] Key points to note during implementation: First, the coordinate system origin calibration must be accurate, using the actual geometric center point of the intersection as the reference. This can be determined through a combination of GPS positioning and manual measurement to avoid errors in subsequent coordinate calculations due to origin offset. It is recommended to review this quarterly to address coordinate updates for scenarios such as road construction and lane marking adjustments. Second, the speed limit (v) must be obtained based on the value indicated by traffic signs at the intersection. If there are variable speed limits at the intersection, such as those adjusted during peak hours, it needs to be linked with the traffic management platform in real time to dynamically update the speed limit parameters and ensure that the sampling frequency (F) is synchronously adapted. Third, panoramic image acquisition must use a high-definition camera, and the shooting angle must be vertical. For road surfaces, to avoid perspective distortion causing measurement errors in the length and width of road markings, image correction algorithms are needed to correct distortion after data acquisition. Fourth, the sampling frequency F needs to be set with upper and lower thresholds. The lower limit should not be lower than 10Hz to avoid sparse sampling leading to trajectory breakage, and the upper limit should not be higher than 30Hz to avoid excessive sampling increasing the computational burden on the device. For special scenarios, such as intersection congestion or low-speed vehicle crawling, a low-frequency sampling mode can be automatically triggered to balance accuracy and efficiency. Fifth, after the coordinate system is established, the coordinate range of key obstacles at the intersection needs to be marked. In subsequent image processing, feature point recognition in such areas can be automatically masked to avoid misjudging them as wheel tracks.

[0021] The image processing module is further expanded and optimized based on the core operations to ensure the accuracy and data integrity of wheel feature point extraction. After the vehicle enters the monitoring range, the camera continuously captures images at a sampling frequency F. First, the YOLO object detection algorithm is used to identify the main body of the vehicle, selecting the complete outline and excluding interference from non-motorized vehicles, pedestrians, and roadside obstacles. At the same time, the lane information of the vehicle is associated to limit the range for subsequent trajectory analysis. When extracting wheel feature points, a combination of edge detection and grayscale thresholding is used to focus on the contact area between the wheel and the road surface. The center point of the contact area is used as the standard feature point. If the contact area is irregular, the center of the smallest bounding rectangle is used to ensure the uniformity of feature point positioning. The coordinates (x, y) of the feature points in each frame and the corresponding timestamp are recorded simultaneously.

[0022] For occlusion scenarios, the position is restored by interpolating the average coordinates of the preceding and following N frames. During implementation, it is important to note that morphological filtering should be used to remove noise such as road surface stains and reflections; defogging and de-reflection preprocessing should be performed in adverse weather conditions; and an early warning should be triggered when there are three or more consecutive occlusions to avoid error accumulation.

[0023] The analysis module sequentially strings together the coordinates of wheel feature points for each frame of the vehicle's passage through the intersection, forming a continuous and complete wheel trajectory curve L. This curve not only includes coordinate data but also synchronously links the timestamps of each node with sampling environment parameters, ensuring the traceability and authenticity of the trajectory. Before comparing with prohibitory traffic markings, the module first calls upon the pre-defined coordinate range library of markings in the environment settings module to clarify the specific coordinate range, marking width, and adjacent lane boundaries of each prohibitory marking in the two-dimensional coordinate system. This establishes a standardized comparison benchmark, preventing the judgment results from being affected by marking positioning deviations.

[0024] The comparison process involves checking frame by frame whether the coordinates of feature points on the trajectory curve L fall within the coordinate range of the prohibitory markings. At the same time, the number and percentage of consecutive frames with overlapping feature points are counted. If feature points overlap with the marking coordinates for 3 or more consecutive frames, or if the number of overlapping frames in a single instance accounts for more than 10% of the total number of frames when a vehicle passes through the intersection, it is judged as a line-crossing behavior. This avoids misjudgment caused by momentary jitter and does not miss continuous line-crossing behavior.

[0025] To determine lane crossing behavior, it is necessary to first clarify the setting standard for the preset distance threshold. Based on the actual width of the intersection markings and in conjunction with lane design specifications, it is recommended that the threshold be set to 0.3 to 0.5 meters. During comparison, the lateral distance of the trajectory curve L exceeding the coordinate range of the markings, i.e., the offset in the x-axis direction, is calculated in real time. If this offset exceeds the preset threshold for multiple consecutive frames, and the maximum lateral distance meets the judgment requirements, and the trajectory as a whole shows a trend of crossing the markings rather than briefly crossing the boundary and immediately returning to the lane, then it is determined to be lane crossing behavior.

[0026] In addition, an anomaly correction mechanism needs to be introduced during the comparison process: if the trajectory curve shows a local abrupt change and the single-frame coordinates deviate significantly from the overall trend, it is necessary to combine the occlusion records of the image processing module with environmental parameters to determine whether it is a feature point extraction error or an occlusion completion deviation. If it is confirmed to be abnormal data, it should be removed to avoid affecting the overall judgment result. At the same time, for special scenarios such as intersection turning and pedestrian avoidance, it is necessary to combine data such as vehicle speed and turning trend to assist in the judgment, eliminate misjudgments caused by legal avoidance, ensure the fairness of the line crossing judgment, and provide a reliable basis for traffic.

[0027] It is worth noting that by adapting the sampling frequency to speed gradient levels, the camera's shooting density can be dynamically adjusted according to the intersection's speed limit. This avoids data redundancy caused by over-sampling of low-speed vehicles while ensuring sufficient sampling frames for high-speed vehicles to fully capture wheel trajectories. This method determines the gradient coefficients using simple mathematical constraints, has clear computational logic, is easy to deploy in embedded systems, and requires no additional hardware costs. It balances data acquisition accuracy with device computing power consumption, adapts to intersection scenarios with different speed limit levels, and provides stable and reliable data support for the subsequent generation of wheel trajectory curves and accurate determination of line violations, thus improving the practicality and adaptability of the entire traffic violation capture system.

[0028] In another preferred embodiment of the present invention, the traffic flow C at the intersection per unit time is obtained; if the traffic flow C is lower than a preset traffic flow threshold C0... min Set the sampling frequency of the intersection camera to 50% of the initial sampling frequency.

[0029] Understandably, the system calculates traffic flow at intersections in real time, based on the number of vehicles detected by cameras and their travel duration. A preset traffic flow threshold is used, which can be calibrated based on historical traffic data and the number of lanes at the intersection. If the detected traffic flow is below the threshold, the intersection is considered to be in a low-load traffic state, and the camera sampling frequency is automatically adjusted to 50% of the initial sampling frequency. If the traffic flow rises above the threshold, the original sampling frequency is restored. This strategy reduces the sampling frequency during low-flow periods, decreasing the amount of data collected by the cameras and the computing power consumed by the edge computing terminals, reducing equipment power consumption and storage pressure, and extending equipment lifespan. Simultaneously, in low-flow scenarios, sparse vehicle traffic eliminates the need for high-density sampling to fully capture wheel trajectories, balancing monitoring accuracy and operational efficiency, and enabling dynamic optimization of system resources.

[0030] In another preferred embodiment of the present invention, if the wheel feature points in the Nth frame before and the Nth frame after are occluded, the wheel feature points in the (N+1)th frame before and the (N+1)th frame after are obtained, and it is determined whether the wheel feature points are occluded. If they are not occluded, the process stops. If they are occluded, the above operation is repeated until the wheel feature points are no longer occluded.

[0031] It is important to note that when occlusion of wheel feature points in the target frame is detected, the system prioritizes retrieving wheel feature point data from the Nth frame before and the Nth frame after. If the feature points within this range are still occluded, a frame range increment strategy is immediately executed to obtain feature point information from the (N+1)th frame before and the (N+1)th frame after, and the occlusion status is reassessed. If valid unoccluded feature points exist in the newly retrieved frame data, the frame range expansion stops, and interpolation calculations are performed based on these valid coordinates to complete the target frame position. If occlusion still exists, the frame range increment operation is repeated until usable feature point data is obtained. This logic effectively avoids trajectory breaks caused by continuous local occlusion, ensuring the integrity of the wheel trajectory curve and providing continuous and reliable data support for subsequent violation determination.

[0032] In another preferred embodiment of the present invention, four wheel trajectory curves corresponding to the vehicle are obtained. When none of the four wheel trajectory curves show any line-crossing or line-exceeding behavior, the vehicle is determined to be driving normally.

[0033] It should be noted that by simultaneously collecting the complete trajectory curves of all four wheels of a vehicle and making a comprehensive judgment, this method overcomes the limitations of traditional single-wheel or dual-wheel trajectory monitoring, significantly improving the accuracy and rigor of violation judgment. It effectively eliminates misjudgments caused by a vehicle's single wheel briefly crossing or over the line without overall lane departure, avoiding disputes due to localized trajectory anomalies and protecting the legitimate rights and interests of vehicle owners; further enhancing the credibility and standardization of the violation capture system.

[0034] In another preferred embodiment of the present invention, when construction barriers, temporary markings, water-covered markings, or snow-buried markings appear at an intersection, the temporary markings are identified from the original prohibitory traffic markings, with the temporary markings serving as the criterion for judgment.

[0035] Understandably, when special circumstances arise at intersections, such as construction barriers, temporary markings, water accumulation covering lines, or snow burying lines, the system can automatically identify temporary markings from the original prohibitory traffic markings, prioritizing the temporary markings as the benchmark for determining violations. This adapts to the dynamic changes at intersections, avoids misjudgments caused by the original markings being invalid or obscured, and ensures the accuracy and rationality of violation determination under special road conditions.

[0036] In another preferred embodiment of the present invention, when the wheel feature points in the image are misaligned, the visible light and infrared light cameras are time-synchronized and calibrated so that the error between the calibrated visible light and infrared light cameras is less than or equal to 1 / F.

[0037] It is worth noting that when wheel feature points are misaligned, the asynchrony between the visible light and infrared cameras can lead to deviations in wheel position data at the same moment, directly affecting the accuracy of the trajectory curve. By synchronizing the calibration to ensure that the error between the two cameras is ≤1 / F, it can be ensured that the two cameras accurately capture wheel feature points within the same sampling period, eliminating the coordinate misalignment problem caused by timing differences.

[0038] This operation ensures the consistency of multi-source image data, providing a reliable data foundation for coordinate completion and trajectory curve generation in occluded scenes, effectively avoiding misjudgments of line pressing and crossing caused by feature point misalignment, and improving the system's judgment accuracy and stability in complex environments.

[0039] In another preferred embodiment of the present invention, if the vehicle length exceeds a preset judgment length, it is classified as a large truck, and the following operations are performed on the large truck: Obtain wheel feature points A1, A2...An on the same side, where n represents the number of wheels on one side. Set a virtual feature point a, and the horizontal and vertical coordinates of the virtual feature point a are the average of the corresponding horizontal and vertical coordinates of wheel feature points A1, A2...An. Generate the wheel trajectory curve LA of the virtual feature point a, and compare the wheel trajectory curve LA with the prohibitory traffic markings.

[0040] It is worth noting that, considering the unique multi-wheel structure of large trucks, virtual feature points are generated by calculating the average coordinates of all wheel feature points on the same side. This constructs a unified trajectory curve for violation comparison. On the one hand, this effectively avoids misjudgments caused by localized deviations of a single wheel, accurately reflecting the overall vehicle trajectory and aligning with the characteristics of large trucks' long bodies and wide wheelbases. On the other hand, simplifying the multi-wheel trajectory into a single virtual trajectory significantly reduces data processing volume and improves system comparison efficiency. Simultaneously, the unified judgment standard takes into account the violation judgment logic of both large trucks and small vehicles, enhancing the system's adaptability to different vehicle types and ensuring the fairness and rigor of the judgment.

[0041] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A video surveillance system based on AI visual inspection, characterized in that, include: Environment Setup Module: Acquire panoramic images of the intersection, establish a two-dimensional coordinate system, where the center point of the intersection is the origin, the horizontal axis represents the transverse direction of the road, and the vertical axis represents the driving direction. Obtain the vehicle speed limit v at the intersection. The method for calculating the sampling frequency F based on the vehicle speed limit v includes: Preset velocity gradient V=λ×v sta v sta Let λ represent the preset unit speed, λ represent the preset gradient coefficient, λ=1, 2..., obtain the value of λ when the constraint condition v / V≤1 is met, and let the sampling frequency F=λ×f, where f represents the initial sampling frequency of the intersection camera; Image processing module: When a vehicle enters the intersection monitoring range, the intersection camera is instructed to take a picture at a sampling frequency F, and the captured image is processed as follows: The vehicle body in the image is obtained, and the wheel feature points in the vehicle body are obtained. The wheel feature points refer to the points where the wheel contacts the road surface. The coordinates (x, y) of the wheel feature points in each frame are recorded. If the wheel feature point is occluded in a single frame image, obtain the coordinates corresponding to the Nth frame before and the Nth frame after. and Let the coordinates of the obscured wheel feature point be... Where N represents the preset offset frame number, N > 0; Analysis module: Records the coordinates of wheel feature points in each frame of the image during the entire process of a vehicle passing through an intersection in chronological order, generates a wheel trajectory curve L, and compares the wheel trajectory curve L with the prohibitory traffic markings. If the wheel track curve L coincides with the prohibitory traffic marking, then the vehicle has crossed the line when passing through the intersection. If the wheel trajectory curve L crosses the prohibitory traffic markings and the maximum lateral distance exceeds a preset distance threshold, then the vehicle has crossed the line when passing through the intersection.

2. The video surveillance system based on AI visual inspection according to claim 1, characterized in that, In the environment setting module, the traffic flow C at the intersection per unit time is obtained. If the traffic flow C is lower than the preset traffic flow threshold C... min Set the sampling frequency of the intersection camera to 50% of the initial sampling frequency.

3. The video surveillance system based on AI visual inspection according to claim 1, characterized in that, In the image processing module, if the wheel feature points in the Nth frame before and the Nth frame after are occluded, the wheel feature points in the (N+1)th frame before and the (N+1)th frame after are obtained, and it is determined whether the wheel feature points are occluded. If they are not occluded, the process stops; if they are occluded, the above operation is repeated until the wheel feature points are no longer occluded.

4. The video surveillance system based on AI visual inspection according to claim 1, characterized in that, In the analysis module, four wheel trajectory curves corresponding to the vehicle are obtained. When none of the four wheel trajectory curves show any line-crossing or line-exceeding behavior, the vehicle is judged to be driving normally.

5. The video surveillance system based on AI visual inspection according to claim 1, characterized in that, In the analysis module, when construction barriers, temporary road markings, road markings covered by water, or road markings buried by snow appear at intersections, the temporary road markings are identified from the original prohibitory traffic markings, with the temporary road markings serving as the criterion for judgment.

6. The video surveillance system based on AI visual inspection according to claim 1, characterized in that, In the analysis module, when the wheel feature points in the image are misaligned, the visible light and infrared cameras are time-synchronized and calibrated so that the error between the calibrated visible light and infrared cameras is less than or equal to 1 / F.

7. The video surveillance system based on AI visual inspection according to claim 1, characterized in that, In the analysis module, if the vehicle length exceeds a preset judgment length, it is classified as a large truck, and the following operations are performed on large trucks: Obtain wheel feature points A1, A2...An on the same side, where n represents the number of wheels on one side. Set a virtual feature point a, and the horizontal and vertical coordinates of the virtual feature point a are the average of the corresponding horizontal and vertical coordinates of wheel feature points A1, A2...An. Generate the wheel trajectory curve LA of the virtual feature point a, and compare the wheel trajectory curve LA with the prohibitory traffic markings.

Images