Road marker detection method based on computer vision
By using ultra-black light cameras and computer vision technology, the displacement and angular deviation of road signs are detected, solving the problems of low detection accuracy and insufficient angular deviation precision in low-light environments, and meeting the safety monitoring needs of high-risk road sections.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGXI JIAOTOU TECHNOLOGY CO LTD
- Filing Date
- 2025-11-07
- Publication Date
- 2026-05-01
AI Technical Summary
Existing technologies have low accuracy (<60%) in detecting road signs in low-light environments and cannot accurately detect angular deviations (error >5°), failing to meet the traffic safety requirements of high-risk road sections.
Video is captured using an ultra-black light camera and decompressed into RGB format frame sequences. Valid frames are selected using an inter-frame difference algorithm and a brightness difference threshold. The transformation matrix is fitted by combining SIFT feature point matching and random sampling consensus algorithm to detect the displacement and angular shift of the marker. Geometric features are detected using Hough transform to improve detection accuracy.
In low-light conditions (0.0002 Lux), the accuracy of marker recognition is improved (≥90%), the positioning error is reduced (≤3cm), the displacement detection sensitivity (≥5cm) and the angle offset detection accuracy (±0.5°) are increased, thus meeting the needs of monitoring the condition of structures in high-risk road sections.
Smart Images
Figure CN121963131A_ABST
Abstract
Description
A computer vision-based method for road sign detection Technical Field
[0001] This invention belongs to the field of computer vision and pattern recognition technology, and specifically relates to a road sign detection method based on computer vision. Background Technology
[0002] High-risk road sections are weak links in highway traffic safety, with accident rates several times higher than those on ordinary road sections. Timely observation of road signs can effectively reduce the accident rate on high-risk road sections. However, due to the combined effects of geological loads, traffic loads, and environmental loads, road signs are prone to positional and angular deviations. Therefore, the detection of positional and angular deviations of road signs is particularly important, as it can promptly address such deviations to reduce the accident rate on high-risk road sections.
[0003] Currently, the detection of road signs in high-risk road sections is generally carried out by cameras. This method has the following problems: low accuracy (<60%) in low light environment (<0.1Lux) and inability to accurately detect angular deviation (error >5°).
[0004] Therefore, a road sign detection method based on computer vision is designed to solve the above problems. Summary of the Invention
[0005] To address the problems mentioned in the background section, this invention provides a computer vision-based road sign detection method. This method improves the accuracy of sign recognition in low-light conditions, reduces positioning errors, enhances displacement detection sensitivity and angular offset detection accuracy, meets the requirements for monitoring the condition of structures in high-risk road sections, and reduces single-frame processing time.
[0006] To achieve the above objectives, the present invention provides the following technical solution: a road sign detection method based on computer vision, comprising the following steps: S1: a super-black light camera acquires video of high-risk road segment structures; S2: the acquired video of high-risk road segment structures is decompressed into an RGB format frame sequence, retaining the brightness channel of the frame image; S3: a preset brightness difference threshold is set, and the brightness difference of consecutive frame images is calculated using an inter-frame difference algorithm. If the brightness difference is less than or equal to the brightness difference threshold, it is determined to be a redundant frame, and only the first frame image is retained; if the brightness difference is greater than the brightness difference threshold, it is determined to be a valid frame, and retained. S4: For the retained frame image, process the luminance channel by limiting contrast adaptive histogram equalization; S5: Reconstruct the processed luminance channel and the original chroma channel of the frame image into an RGB format image; S6: Acquire multi-angle images of road signs under normal lighting, extract SIFT feature points, and store them as a template library; S7: Extract SIFT feature points from the reconstructed RGB format image, perform K-nearest neighbor matching with SIFT feature points in the template library, and set a preset matching point threshold. If the number of matching points does not exceed the preset matching point threshold, retry; if the number of matching points exceeds the preset matching point threshold... The transformation matrix is fitted using a random sampling consensus algorithm. SIFT feature points matched by the template library are used to locate the bounding rectangle region of the marker in the RGB format image using the transformation matrix. S8: Obtain the geometric center coordinates of the bounding rectangle region of the located marker in the RGB format image. Convert them to physical coordinates based on the intrinsic parameters of the ultra-black light camera. Calculate the displacement of the physical coordinates between consecutive frames of RGB format images. Set a preset displacement anomaly threshold. If the displacement of the physical coordinates between two consecutive frames of RGB format images exceeds the preset displacement anomaly threshold, it is marked as an abnormal frame of RGB format image. Simultaneously, store the physical coordinates and timestamp. S9: Acquire multiple frames of RGB format images of the marker in a stable state. Detect the geometric features of the marker using Hough transform. Fit the initial rotation angle of the marker as a reference value. Extract the long edge line of the bounding rectangle region of the located marker in the RGB format image. Calculate the angle corresponding to the slope of the line as the current rotation angle. Calculate the difference between the current rotation angle and the reference value as the angle offset. Set a preset angle offset anomaly threshold. If the angle offset exceeds the preset angle offset anomaly threshold, it is marked as an abnormal frame of RGB format image. Simultaneously, store the physical coordinates and timestamp.
[0007] Further, in step S3, the specific steps for calculating the brightness difference of consecutive frame images using the inter-frame difference algorithm include: acquiring two consecutive frame images; converting the two consecutive frame images into grayscale images respectively; performing pixel-by-pixel absolute value difference on the two consecutive grayscale images to obtain a difference image; setting a brightness difference threshold, and binarizing the difference image by converting pixel values less than or equal to the brightness difference threshold to one and pixel values greater than the brightness difference threshold to zero; counting the total number of pixels with a value of 1 in the binarized difference image to determine the brightness difference between the two consecutive frame images.
[0008] Further, in step S4, the specific steps of processing the luminance channel by limiting contrast adaptive histogram equalization include: converting the retained frame image to the YUV color space and extracting the luminance channel; uniformly dividing the luminance channel image into multiple 8×8 pixel non-overlapping sub-blocks; statistically analyzing the histogram of the pixel grayscale values of each 8×8 pixel non-overlapping sub-block, calculating the number of pixels at each grayscale level, calculating the maximum number of pixels allowed at each grayscale level according to a preset clipping limit, and for grayscale levels where the number of pixels in the histogram exceeds the maximum number of pixels allowed at the grayscale level, uniformly distributing the excess pixels to all grayscale levels to complete the histogram clipping; equalizing the clipped histogram to obtain the transformation function of the pixel grayscale values within the sub-block, mapping the sub-block pixels to a new grayscale range, and enhancing the local contrast of the sub-block; the final grayscale values of adjacent sub-block pixels are calculated by bilinear interpolation from the equalization output of the adjacent sub-blocks.
[0009] Further, in step S7, the specific steps of reconstructing the RGB format image and SIFT feature points in the template library by K-nearest neighbor matching include: calculating the Euclidean distance between the reconstructed RGB format image and the SIFT feature points in the template library; sorting the distances between each SIFT feature point in the reconstructed RGB format image and the SIFT feature points in the template library; and selecting the SIFT feature points in the reconstructed RGB format image and the SIFT feature points in the template library with the smallest two distances as matching points.
[0010] Further, in step S7, the specific steps of fitting the transformation matrix using the random sampling consensus algorithm include: randomly selecting multiple sets of matching points, converting the homography transformation of the selected sets of matching points into a system of linear equations, solving the system of linear equations using the least squares method to obtain an initial homography matrix; verifying all matching points using the initial homography matrix, calculating the reprojection error of each set of matching points; setting a preset error threshold, if the reprojection error of each set of matching points does not exceed the error threshold, then the set of matching points is taken as interior points; counting the number of interior points; repeating the above steps multiple times, each time recording the homography matrix with the most interior points and the corresponding set of interior points; finally, fitting the homography matrix with the most interior points using the least squares method to obtain the transformation matrix.
[0011] Further, in step S9, the specific steps for detecting the geometric features of the marker using the Hough transform include: converting the acquired multi-frame RGB format images into grayscale images; extracting the edges of the grayscale images using the Canny operator to obtain a binarized edge image; parametrically defining the straight line of the binarized edge image using polar coordinates, including the perpendicular distance from the straight line to the origin of the binarized edge image and the angle between the straight line and the X-axis of the binarized edge image; discretizing the angle between the straight line and the X-axis of the binarized edge image into several sampling points; and calculating the edge point corresponding to each discrete angle. Extreme radius; Maintain a two-dimensional accumulator array. For each pair of extreme radii calculated at each edge point, perform a voting operation on the angle between the line and the X-axis of the binarized edge image. Preset a voting threshold. If the two-dimensional accumulator array exceeds the preset voting threshold, the extreme radius and the angle between the line and the X-axis of the binarized edge image are identified as a salient line in the image. Calculate the opposite side length and the adjacent side angle difference of all identified salient lines. If the opposite side lengths of the four salient lines are equal and the adjacent side angle difference is 90°, then the four salient lines are determined to be the geometric features of the marker.
[0012] Further, in step S9, the specific steps for fitting the initial rotation angle of the marker include: obtaining the endpoint coordinates of the four straight lines of the marker's geometric features; calculating the lengths of the four straight lines of the marker's geometric features based on the endpoint coordinates; determining the long side of the marker's geometric features based on the lengths of the four straight lines; using the angle of the long side group of the marker's geometric features as the initial angle, converting it to the range of -180° to 180° to obtain the initial rotation angle.
[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The present invention acquires video of high-risk road structure by using an ultra-black light camera, decompresses the acquired video of high-risk road structure into a frame sequence in RGB format, retains the brightness channel of the frame image, and reassembles the processed brightness channel with the original color channel of the frame image by limiting the contrast adaptive histogram equalization to form an RGB format image. This can improve the accuracy of marker recognition (≥90%) in low-light environment (0.0002Lux) and reduce positioning error (≤3cm).
[0014] 2. This invention constructs a template library, matches SIFT feature points to locate the bounding rectangle region of the marker in the RGB format image, calculates the displacement and angular offset of the bounding rectangle region of consecutive frames of RGB format images, and compares it with preset displacement anomaly thresholds and angular offset anomaly thresholds to determine the position and angular offset of road markers in high-risk road sections. This can improve the displacement detection sensitivity (≥5cm) and angular offset detection accuracy (±0.5°), meeting the needs of monitoring the status of structures in high-risk road sections.
[0015] 3. The present invention presets a brightness difference threshold and calculates the brightness difference of consecutive frame images through an inter-frame difference algorithm. If the brightness difference is less than or equal to the brightness difference threshold, it is determined to be a redundant frame and only the first frame image is retained. If the brightness difference is greater than the brightness difference threshold, it is determined to be a valid frame and retained. Only the retained valid frames are processed, which improves the processing time of a single frame (≤100ms). Attached Figure Description
[0016] Figure 1 is a flowchart of the method of 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] This invention provides the following technical solution: a road sign detection method based on computer vision, comprising the following steps: S1: A super-black light camera acquires video of high-risk road segment structures; S2: The acquired video of high-risk road segment structures is decompressed into an RGB format frame sequence, retaining the brightness channel of the frame image; S3: A preset brightness difference threshold is set, and the brightness difference of consecutive frame images is calculated using an inter-frame difference algorithm. If the brightness difference is less than or equal to the brightness difference threshold, it is determined to be a redundant frame, and only the first frame image is retained; if the brightness difference is greater than the brightness difference threshold, it is determined to be a valid frame and retained; S4: The retained... The frame image is processed by limiting contrast and adaptive histogram equalization for the luminance channel; S5: The processed luminance channel is reconstructed into an RGB format image with the original chroma channel of the frame image; S6: Multi-angle images of road signs under normal lighting are acquired, SIFT feature points are extracted, and stored as a template library; S7: SIFT feature points of the reconstructed RGB format image are extracted and K-nearest neighbor matching is performed with SIFT feature points in the template library. A preset matching point threshold is used. If the number of matching points does not exceed the preset matching point threshold, the process is retried. If the number of matching points exceeds the preset matching point threshold, the process is passed. The random sampling consensus algorithm fits the transformation matrix, and the SIFT feature points matched by the template library are used to locate the bounding rectangle region of the marker in the RGB format image through the transformation matrix; S8: Obtain the geometric center coordinates of the bounding rectangle region of the located marker in the RGB format image, convert them to physical coordinates based on the intrinsic parameters of the ultra-black light camera, calculate the displacement of the physical coordinates between consecutive frames of RGB format images, and set a preset displacement anomaly threshold. If the displacement of the physical coordinates between two consecutive frames of RGB format images exceeds the preset displacement anomaly threshold, it is marked as an abnormal frame of RGB format image, and the physical coordinates and timestamp are stored at the same time; S9: Collect multiple frames of RGB format images of the marker in a stable state, detect the geometric features of the marker through Hough transform, fit the initial rotation angle of the marker as the reference value, extract the long edge line of the bounding rectangle region of the located marker in the RGB format image, calculate the angle corresponding to the slope of the line as the current rotation angle, calculate the difference between the current rotation angle and the reference value as the angle offset, set a preset angle offset anomaly threshold. If the angle offset exceeds the preset angle offset anomaly threshold, it is marked as an abnormal frame of RGB format image, and the physical coordinates and timestamp are stored at the same time.
[0019] Specifically, in step S3, the specific steps for calculating the brightness difference of consecutive frame images using the inter-frame difference algorithm include: acquiring two consecutive frame images; converting the two consecutive frame images into grayscale images respectively; performing pixel-by-pixel absolute value difference on the two consecutive grayscale images to obtain a difference image; setting a preset brightness difference threshold, and binarizing the difference image by converting pixel values less than or equal to the brightness difference threshold to one and pixel values greater than the brightness difference threshold to zero; counting the total number of pixels with a value of 1 in the binarized difference image to determine the brightness difference between the two consecutive frame images.
[0020] Specifically, in step S4, the specific steps of processing the luminance channel through contrast-adaptive histogram equalization include: converting the retained frame image to the YUV color space and extracting the luminance channel; uniformly dividing the luminance channel image into multiple 8×8 pixel non-overlapping sub-blocks, wherein if the edge size of the luminance channel image is less than 8×8, zero-padding is applied to the edge region; calculating the histogram of the pixel grayscale values of each 8×8 pixel non-overlapping sub-block, calculating the number of pixels at each grayscale level, calculating the maximum number of pixels allowed at each grayscale level according to the preset clipping limit, and for grayscale levels where the number of pixels in the histogram exceeds the maximum number of pixels allowed at the grayscale level, uniformly distributing the excess pixels to all grayscale levels to complete the histogram clipping; the expression for calculating the maximum number of pixels allowed at each grayscale level is: In the formula: Indicates the preset clipping limit; Indicates the total number of pixels in the sub-block; This represents the total number of gray levels; equalization is applied to the clipped histogram to obtain a transformation function for the gray values of pixels within the sub-block, mapping the sub-block pixels to a new gray range and enhancing the local contrast of the sub-block; the expression for the transformation function is: In the formula: Indicates grayscale level; This indicates that the grayscale level within the edited sub-block is... The probability of a pixel appearing at a given time; the expression for mapping sub-block pixels to a new grayscale range is: In the formula: This indicates rounding to the nearest integer. This represents the transformation function; the final grayscale value of adjacent sub-block pixels is calculated by bilinear interpolation from the equalization output of the adjacent sub-blocks.
[0021] Specifically, in step S7, the steps for reconstructing the RGB format image and SIFT feature points in the template library using K-nearest neighbor matching include: calculating the Euclidean distance between the reconstructed RGB format image and the SIFT feature points in the template library, expressed as: In the formula: These represent SIFT feature points used to reconstruct an RGB format image. Represents SIFT feature points in the template library; Indicates the number of nearest neighbors. For each SIFT feature point in the reconstructed RGB format image, sort its distances with the SIFT feature points in the template library; select the SIFT feature points in the reconstructed RGB format image with the two smallest distances and the SIFT feature points in the template library as matching points.
[0022] Specifically, in step S7, the steps for fitting the transformation matrix using the random sampling consensus algorithm include: randomly selecting multiple sets of matching points, transforming the homography transformation of the selected sets of matching points into a system of linear equations, solving the system of linear equations using the least squares method to obtain the initial homography matrix; verifying all matching points using the initial homography matrix, and calculating the reprojection error of each set of matching points, expressed as: In the formula: Indicates the actual physical coordinates of the matching point; Indicates the predicted physical coordinates of the matching point; The scaling factor for predicting physical coordinates is used; a preset error threshold is set, and if the reprojection error of each set of matching points does not exceed the error threshold, then the set of matching points is taken as interior points; the number of interior points is counted; the above steps are repeated multiple times, and each time the homography matrix with the most interior points and the corresponding set of interior points are recorded; finally, the interior point with the most interior points is fitted to the homography matrix using the least squares method to obtain the transformation matrix.
[0023] Specifically, in step S8, the specific steps for converting the geometric center coordinates to physical coordinates based on the intrinsic parameters of the ultra-black light camera include: the conversion relationship between the geometric center coordinates and physical coordinates is as follows: In the formula: This indicates the vertical distance from the ultra-black light camera to the plane where the marker is located; This represents the pixel coordinates of the matching point in the image pixel coordinate system; Indicates the pixel coordinates of the principal point; This indicates the focal length of an ultra-black light camera.
[0024] Specifically, in step S9, the specific steps for detecting the geometric features of the marker using the Hough transform include: converting the acquired multi-frame RGB format images into grayscale images; extracting the edges of the grayscale images using the Canny operator to obtain a binarized edge image; and parameterizing the straight line of the binarized edge image using polar coordinates, including the perpendicular distance from the straight line to the origin of the binarized edge image and the angle between the straight line and the X-axis of the binarized edge image, expressed as: In the formula: This represents the perpendicular distance from the line to the origin of the coordinate system. Indicates edge points; Indicates vertical distance and The angle between the line and the X-axis of the binarized edge image is discretized into several sampling points; the polar radius corresponding to each discrete angle edge point is calculated, expressed as: In the formula: Indicates edge points; Indicates the perpendicular distance of a line and its direction. The included angle of the axes; maintain a two-dimensional accumulator array. For each edge point, calculate each pair Perform voting operations; preset voting threshold, if the two-dimensional accumulator array If the preset voting threshold is exceeded, then... Identify a prominent line in the image; calculate the length of the opposite side and the angle difference between the adjacent sides of all identified prominent lines. If the length of the opposite side of the four prominent lines is equal and the angle difference between the adjacent sides is 90°, then the four prominent lines are determined to be the geometric features of the marker.
[0025] Specifically, in step S9, the steps for fitting the initial rotation angle of the marker include: obtaining the endpoint coordinates of the four straight lines of the marker's geometric features; calculating the lengths of the four straight lines of the marker's geometric features based on the endpoint coordinates; and determining the long side of the marker's geometric features based on the lengths of the four straight lines, expressed as: In the formula: and Represents the coordinates of the two endpoints of the straight line representing the geometric feature of the marker; using the angle of the long side group of the geometric feature of the marker as the initial angle, transforming it to the range of -180° to 180°, we obtain the initial rotation angle, expressed as: .
[0026] Specifically, in step S9, the expression for calculating the angle corresponding to the slope of the straight line is: In the formula: Indicates the slope of a straight line; In the formula: Indicates the perpendicular distance of a line and its direction. The included angle of the axis.
[0027] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A road sign detection method based on computer vision, characterized in that, Includes the following steps: S1: Collect video footage of structures in high-risk road sections; S2: Decompress the video of high-risk road structure into a frame sequence, retaining the brightness channel of the frame image; S3: Preset brightness difference threshold, calculate the brightness difference of consecutive frame images. If the brightness difference is less than or equal to the brightness difference threshold, it is determined to be a redundant frame and only the first frame image is retained. Otherwise, it is determined to be a valid frame and is retained. S4: Process the luminance channel of the retained frame image; S5: Reassemble the processed luminance channel and the original chrominance channel into an image; S6: Acquire multi-angle images of road signs under normal lighting, extract SIFT feature points, and store them as a template library. S7: Extract SIFT feature points from the reconstructed image and match them with SIFT feature points in the template library. Set a preset matching point threshold. If the number of matching points does not exceed the preset matching point threshold, retry. Otherwise, fit a transformation matrix. The template library matches SIFT feature points and uses the transformation matrix to locate the bounding rectangle region of the marker in the image. S8: Obtain the geometric center coordinates of the bounding rectangle region of the located marker in the image, convert them into physical coordinates, calculate the physical coordinate displacement between consecutive frames, preset anomaly threshold, if the physical coordinate displacement between two consecutive frames exceeds the preset anomaly threshold, mark the abnormal frame image, and store the physical coordinates and timestamp at the same time. S9: Acquire multiple frames of images of the marker in a stable state, detect the geometric features of the marker, fit the initial rotation angle of the marker as a reference value, extract the long edge line of the bounding rectangle region of the marker in the image, calculate the angle corresponding to the slope of the line as the current rotation angle, calculate the difference between the current rotation angle and the reference value as the angle offset, preset the angle offset abnormal threshold, if the angle offset exceeds the preset angle offset abnormal threshold, mark the abnormal frame image, and store the physical coordinates and timestamp.
2. The road sign detection method based on computer vision according to claim 1, characterized in that: In step S3, the specific steps for calculating the brightness difference between consecutive frames include: acquiring two consecutive frames; converting the two consecutive frames into grayscale images respectively; performing pixel-by-pixel absolute value difference on the two consecutive grayscale images to obtain a difference image; setting a brightness difference threshold, and binarizing the difference image by converting pixel values less than or equal to the brightness difference threshold to one and pixel values greater than the brightness difference threshold to zero; counting the total number of pixels with a value of 1 in the binarized difference image to determine the brightness difference between the two consecutive frames.
3. The road sign detection method based on computer vision according to claim 2, characterized in that: In step S4, the specific steps for processing the luminance channel include: converting the retained frame image to the YUV color space and extracting the luminance channel; uniformly dividing the luminance channel image into multiple 8×8 pixel non-overlapping sub-blocks; statistically analyzing the histogram of the pixel grayscale values of each 8×8 pixel non-overlapping sub-block, calculating the number of pixels at each grayscale level, calculating the maximum allowed number of pixels at each grayscale level according to a preset clipping limit, and for grayscale levels where the number of pixels in the histogram exceeds the maximum allowed number of pixels, uniformly distributing the excess pixels to all grayscale levels to complete histogram clipping; equalizing the clipped histogram to obtain the transformation function of the pixel grayscale values within the sub-block, mapping the sub-block pixels to a new grayscale range, and enhancing the local contrast of the sub-block; and calculating the final grayscale values of adjacent sub-block pixels by bilinear interpolation from the equalization output of the adjacent sub-blocks.
4. The road sign detection method based on computer vision according to claim 3, characterized in that: In step S7, the specific steps for matching SIFT feature points in the reconstructed image and the template library include: calculating the Euclidean distance between the SIFT feature points in the reconstructed image and the template library; sorting the SIFT feature points of each reconstructed image by their distances to the SIFT feature points in the template library; and selecting the SIFT feature points of the reconstructed image and the SIFT feature points in the template library with the two smallest distances as matching points.
5. The road sign detection method based on computer vision according to claim 4, characterized in that: In step S7, the specific steps for fitting the transformation matrix include: randomly selecting multiple sets of matching points, transforming the homography transformation of the selected sets of matching points into a system of linear equations, solving the system of linear equations using the least squares method to obtain an initial homography matrix; verifying all matching points using the initial homography matrix, calculating the reprojection error of each set of matching points; setting a preset error threshold, if the reprojection error of each set of matching points does not exceed the error threshold, then the set of matching points is taken as interior points; counting the number of interior points; repeating the above steps multiple times, recording the homography matrix with the most interior points and the corresponding set of interior points each time; finally, fitting the homography matrix with the most interior points using the least squares method to obtain the transformation matrix.
6. The road sign detection method based on computer vision according to claim 5, characterized in that: In step S9, the specific steps for detecting the geometric features of the marker include: converting the acquired multi-frame RGB format images into grayscale images; extracting the edges of the grayscale images using the Canny operator to obtain a binarized edge image; parametrically representing the straight lines of the binarized edge image using polar coordinates, including the vertical distance from the straight line to the origin of the binarized edge image and the angle between the straight line and the X-axis of the binarized edge image; discretizing the angle between the straight line and the X-axis of the binarized edge image into several sampling points; calculating the polar radius corresponding to each discrete angle edge point; maintaining a two-dimensional accumulator array, and performing a voting operation for each pair of polar radii calculated for each edge point and the angle between the straight line and the X-axis of the binarized edge image; setting a preset voting threshold, and if the two-dimensional accumulator array exceeds the preset voting threshold, then the polar radius and the angle between the straight line and the X-axis of the binarized edge image are identified as a significant straight line in the image; calculating the opposite side length and the angle difference between adjacent sides of all identified significant straight lines, and if the opposite side lengths of the four significant straight lines are equal and the angle difference between adjacent sides is 90°, then the four significant straight lines are determined to be geometric features of the marker.
7. The road sign detection method based on computer vision according to claim 6, characterized in that: In step S9, the specific steps for fitting the initial rotation angle of the marker include: obtaining the endpoint coordinates of the four straight lines of the marker's geometric features; calculating the lengths of the four straight lines of the marker's geometric features based on the endpoint coordinates; determining the long side of the marker's geometric features based on the lengths of the four straight lines; using the angle of the long side group of the marker's geometric features as the initial angle, converting it to the range of -180° to 180° to obtain the initial rotation angle.