Robot navigation method based on vision

Abstract

The invention relates to the technical field of visual navigation, in particular to a robot navigation method based on vision, which comprises the following steps: acquiring a grayscale image to extract dynamic point locations, screening boundary features to reconstruct a feature set, calculating gradient change to generate candidate guide points, converting coordinates to construct a path track set, and outputting a navigation planning path. According to the method, rapid elimination of static backgrounds is realized through difference threshold selection, dynamic region extraction accuracy is effectively improved, combined screening of boundary line feature point frequency and average feature density is combined, key feature stability is enhanced, and the identification capability of image local structure mutation is enhanced by using a variance change rate of a gradient amplitude sequence. Robustness of path guide point extraction is improved through judgment of an abnormal section, dynamic construction of a path track is realized by means of continuous time integration of image coordinates, path identification precision and spatial positioning continuity are optimized, and continuity, precision and real-time performance of navigation path planning are improved.

Description

Technical Field

[0001] The present invention relates to the field of visual navigation technology, and in particular to a vision-based robot navigation method. Background Art

[0002] The field of visual navigation technology involves control technology that guides mobile devices or systems to perform path identification and position judgment by acquiring and analyzing environmental image information. Its core issues include image acquisition, feature extraction, environmental mapping, path identification, and pose estimation. It is an important basic technology for realizing autonomous navigation and environmental perception in scenarios such as robots, unmanned driving, and intelligent manufacturing. Combining computer vision and motion control theory, it is committed to improving navigation accuracy and environmental adaptability. Among them, the traditional vision-based robot navigation method refers to using a camera to collect image information of the surrounding environment, and obtaining the current position and orientation of the robot through image feature matching and pose inference. It relies on visual odometry or image sequences to realize navigation path planning and position update. It usually uses feature point extraction and matching algorithms combined with inter-frame motion estimation to complete the target tracking and path guidance process.

[0003] In traditional navigation methods, the navigation path update is mainly completed by feature point matching combined with inter-frame motion estimation. Under conditions of drastic lighting changes or significant background interference, it is easy to cause unstable feature extraction. Especially in dynamic environments, the image background and foreground are highly integrated, and the feature point positioning accuracy decreases, which in turn affects the accuracy of pose estimation. In addition, relying on fixed feature matching criteria lacks the ability to sensitively capture image mutation areas, resulting in significant deviations in the extracted position of path guide points, causing the navigation path to be broken or offset. For example, in low-texture areas or repetitive pattern backgrounds, path tracking failures and positioning drift are prone to occur, affecting the overall stability and adaptability of the navigation system. Summary of the Invention

[0004] The purpose of the present invention is to solve the shortcomings of the prior art and to propose a robot navigation method based on vision.

[0005] In order to achieve the above object, the present invention adopts the following technical solution: a vision-based robot navigation method, comprising the following steps:

[0006] S1: Obtain the continuous grayscale images captured by the robot's forward camera, detect the grayscale difference value of each pixel in the image at the same coordinate, mark the static background feature points and remove the corresponding coordinates, and generate the number of dynamic grayscale change points;

[0007] S2: Based on the image coordinate set corresponding to the number of dynamic grayscale change points, combined with the edge detection boundary, the frequency of occurrence of boundary feature points is detected, the average feature density within the boundary is calculated, the segments below the feature matching threshold are screened out and the feature set is reconstructed to generate boundary feature retention data;

[0008] S3: according to the image frame corresponding to the boundary feature retention data, set a horizontal scanning band, extract the gradient amplitude sequence, calculate the variance change rate, and generate gradient change amplitude distribution data;

[0009] S4: Based on the gradient change amplitude distribution data, determine the column segments in the sequence whose continuous gradient values ​​are lower than the inlier threshold, select the column coordinates corresponding to the position with the most drastic grayscale difference change as the candidate path guide point, and generate a candidate guide point position record;

[0010] S5: performing coordinate conversion according to the image coordinates in the candidate guide point position record, integrating all converted coordinate points in chronological order, constructing a path trajectory set, and obtaining a visual robot navigation planning path.

[0011] As a further solution of the present invention, the number of dynamic grayscale change points includes the distribution of image active areas, pixel change intensity statistics, and initial contour information of motion boundaries; the boundary feature retention data includes local structure density distribution, edge feature stability description, and candidate navigation element screening results; the gradient change amplitude distribution data includes image texture clarity layering information, edge response intensity comparison data, and image structure mutation trend information; the candidate guide point position records include a candidate set of navigable channel coordinates, image path guidance key point distribution, and spatial consistency verification reference values; the visual robot navigation planning path includes a time series path point set, an environment space model mapping result, and a dynamic path planning reference trajectory.

[0012] As a further solution of the present invention, the specific steps for obtaining the number of dynamic grayscale change points are:

[0013] S111: Obtain a continuous grayscale image sequence captured by the robot's forward camera, detect the grayscale values ​​of the same coordinate position in the image between adjacent frames, and calculate the grayscale difference of each pixel between consecutive frames. Based on the grayscale difference and the coordinate index, a corresponding mapping is formed to obtain a pixel grayscale difference mapping result;

[0014] S112: Based on the pixel grayscale difference mapping result, perform OTSU adaptive threshold judgment on all pixel grayscale differences according to the principle of maximizing inter-class variance, filter out pixels below the threshold, determine the corresponding coordinates as static background feature points, and eliminate the pixel coordinates to obtain a coordinate set of the dynamically changing area;

[0015] S113: performing point-by-point statistics on the pixel difference values ​​of each coordinate in the reserved area according to the dynamically changing area coordinate set, summarizing the pixel difference data and combining the coordinate quantity to form a comparison quantity, and obtaining the number of dynamic grayscale change points.

[0016] As a further solution of the present invention, the specific steps for obtaining the boundary feature preservation data are:

[0017] S211: Based on the image coordinate set corresponding to the number of dynamic grayscale change points, perform an edge extraction operation on the original image using Canny edge detection, extract boundary segments and mark boundary pixels, count the frequency of feature points on each boundary segment and map them to boundary coordinate positions to obtain boundary feature occurrence frequency data;

[0018] S212: Based on the boundary feature occurrence frequency data, calculate the gradient direction change value of each boundary segment, determine the amplitude distribution, construct a directional histogram, count the frequency density corresponding to the directional distribution, summarize the directional gradient histograms on each boundary segment and take the average to generate boundary directional feature density data;

[0019] S213: Based on the boundary directional feature density data and combined with the ORB feature point matching judgment standard, the difference between the directional feature value of each boundary segment and the overall feature mean of the image is judged, the deviated area is screened out and the remaining boundary segment data is retained, the coordinates and feature information sets of the retained boundary segments are reorganized, a unified feature data index is established, and the boundary feature retention data is obtained.

[0020] As a further solution of the present invention, the specific steps for obtaining the gradient change amplitude distribution data are:

[0021] S311: According to the image frame corresponding to the boundary feature retention data, set the ROI horizontal scanning band, extract the grayscale value distribution of all pixels in the area in each scanning band according to the image column direction, convolve each pixel in the image coordinate area to extract the edge gradient response, obtain the pixel with the maximum Sobel gradient value in each column and record the amplitude value, and obtain the scanning band gradient amplitude sequence group;

[0022] S312: Based on the scanning band gradient amplitude sequence group, performing difference calculation on adjacent column values ​​in each sequence, constructing a sliding window and calculating the approximate second-order grayscale derivative of the corresponding pixel, performing Laplacian response calculation on each column, and synchronously performing variance calculation on the obtained Laplacian values. The variance results of all sliding windows are mean-normalized and sorted by column index to obtain a Laplacian variance sequence;

[0023] S313: Based on the Laplacian variance sequence, perform differential analysis on the variance values ​​between two adjacent columns, calculate the offset between the absolute value of all differences and the average difference, normalize all offsets, and perform comprehensive weighting based on the Sobel amplitude change between columns. Then, sort the change rate of each column, calculate the gradient change response value, and establish a difference intensity distribution map.

[0024] As a further solution of the present invention, the specific steps for obtaining the candidate guidance point position record are:

[0025] S411: Based on the gradient change amplitude distribution data, read the gradient amplitude value of each column one by one, and continuously slide according to the column index to determine whether it is lower than the RANSAC inlier threshold, record the column index sequence that continuously satisfies the gradient amplitude value lower than the threshold, record the column segment length and establish the corresponding index structure, and obtain the continuous low-amplitude column segment length set;

[0026] S412: According to the set of continuous low-amplitude column segment lengths, Hough transform line segment detection parameters are set to determine the length of each column segment, and the start and end column coordinates and lengths of all column segments that meet the conditions are recorded one by one. The grayscale difference sequence within the range is extracted to obtain a candidate line segment grayscale difference sequence group.

[0027] S413: Based on the candidate line segment grayscale difference sequence group, linkage scoring is performed according to the column segment sequence length, grayscale difference change intensity and Sobel gradient change value, and the candidate column significance score value is obtained by calculation. Sorting is performed within the column segment according to the score value, and the column index with the maximum significance is determined as the target guide point of the current segment, and the candidate guide point position record is generated.

[0028] As a further solution of the present invention, the specific steps for obtaining the visual robot navigation planning path are:

[0029] S511: Extracting the pixel positions of the marked guide points in each frame based on the image coordinates in the candidate guide point position record, and setting a world coordinate reference template accordingly, constructing a corresponding point set between the image coordinates and the world coordinates, mapping the positions of all image guide points in the image coordinate system to the three-dimensional world coordinate system, and generating a camera calibration transformation coordinate set;

[0030] S512: Based on the camera calibration transformation coordinate set, reorder the image frames to which the guide points belong according to the time sequence, extract the three-dimensional coordinate points in adjacent time frames, calculate the position interval and direction change segment by segment, and classify the spatial position difference information between each pair of coordinate points into a set of path segment description data. All path segments are sequentially connected to form an ordered spatial path data structure to obtain a path trajectory set;

[0031] S513: Based on the three-dimensional point set structure of all continuous segments in the path trajectory set, the connection relationship between the path segments is screened, and the path segments with continuous time intervals and spatial distance changes within the set range are marked as connectable segments. The inter-segment connection processing is performed in sequence, and the coordinate point sequences of all continuous segments are integrated to establish the visual robot navigation planning path.

[0032] Compared with the prior art, the advantages and positive effects of the present invention are:

[0033] In the present invention, the static background is quickly eliminated through difference value threshold selection, which effectively improves the accuracy of dynamic area extraction. The frequency of boundary line feature points and the average feature density are combined to screen the stability of key features, and the variance change rate of the gradient amplitude sequence is used to enhance the recognition ability of local structural mutations in the image. The robustness of path guide point extraction is improved by judging abnormal sections, and the dynamic construction of path trajectory is realized with the help of continuous time integration of image coordinates, which optimizes the path recognition accuracy and continuity of spatial positioning, enhances the adaptability to image disturbances and the response capability under environmental structure changes during navigation execution, and improves the continuity, accuracy and real-time performance of navigation path planning. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flow chart of the main steps of the present invention;

[0035] Figure 2 This is a flow chart for obtaining the number of dynamic grayscale change points in the present invention;

[0036] Figure 3 A flow chart for obtaining boundary feature retention data of the present invention;

[0037] Figure 4 This is a flow chart for obtaining gradient change amplitude distribution data of the present invention;

[0038] Figure 5 A flowchart for obtaining candidate guidance point location records of the present invention;

[0039] Figure 6 This is a flow chart for obtaining the path planning for the visual robot navigation of the present invention. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0041] In the description of the present invention, it should be understood that the terms "length," "width," "up," "down," "front," "back," "left," "right," "vertical," "horizontal," "top," "bottom," "inside," "outside," and the like, indicating positions or relationships, are based on the positions or relationships shown in the accompanying drawings and are intended only to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limiting the present invention. Furthermore, in the description of the present invention, "plurality" means two or more, unless otherwise expressly and specifically defined.

[0042] See also Figure 1 , a vision-based robot navigation method, comprising the following steps:

[0043] S1: Obtain continuous grayscale images captured by the robot's forward camera, detect the grayscale difference value of each pixel in the image at the same coordinate, perform pixel-by-pixel judgment on the difference value based on the OTSU adaptive threshold (an adaptive binarization method based on maximizing inter-class variance), mark pixels below the threshold as static background feature points, and remove the corresponding coordinates. Only the coordinates of the area where the pixel difference is greater than the threshold are retained to generate the number of dynamic grayscale change points;

[0044] S2: Based on the image coordinate set corresponding to the number of dynamic grayscale change points, combined with the Canny edge detection (multi-stage edge detection standard) boundary, the frequency of feature points within the unit pixel length on each boundary line is detected, and the average feature density of the HOG feature (Histogram of Directed Gradients (proposed by Dalal, used for pedestrian detection)) within the boundary is calculated. The segments below the ORB feature matching threshold (feature matching standard) are screened out and the feature set is reconstructed to generate boundary feature retention data;

[0045] S3: Based on the boundary feature retention data corresponding to the image frame, set the ROI (region of interest) horizontal scanning band, extract the Sobel operator (discrete differential operator (3×3 convolution kernel)) gradient amplitude sequence in the scanning band, calculate the Laplacian variance (image clarity evaluation index (larger values ​​indicate clearer edges)) change rate of continuous columns in the sequence, and generate gradient change amplitude distribution data;

[0046] S4: Based on the gradient change amplitude distribution data, determine the column segments in the sequence whose continuous gradient values ​​are lower than the RANSAC inlier threshold (the inlier threshold of the random sampling consensus algorithm (usually 0.5-1.5 pixels)), record the length of the continuous column segment, and compare it with the Hough transform line segment detection parameters (ρ accuracy (1 pixel) and θ accuracy (1 degree) for line detection), select the column coordinates corresponding to the position with the most drastic change in grayscale difference as the candidate path guide point, and generate the candidate guide point position record;

[0047] S5: Based on the image coordinates in the candidate guidance point position record, perform the Zhang camera calibration method (a standard camera calibration method based on a planar calibration plate) coordinate transformation, integrate all the transformed coordinate points in chronological order, construct a path trajectory set, and obtain the visual robot navigation planning path.

[0048] The number of dynamic grayscale change points includes the distribution of image active areas, pixel change intensity statistics, and initial contour information of motion boundaries. The boundary feature retention data includes the local structure density distribution, edge feature stability description, and candidate navigation element screening results. The gradient change amplitude distribution data includes image texture clarity layering information, edge response intensity comparison data, and image structure mutation trend information. The candidate guidance point position records include the candidate set of navigable channel coordinates, the distribution of image path guidance key points, and the spatial consistency verification reference value. The visual robot navigation planning path includes the time series path point set, the environment space model mapping result, and the dynamic path planning reference trajectory.

[0049] See also Figure 2 , the specific steps of S1 are:

[0050] S111: Obtain a continuous grayscale image sequence captured by the robot's forward camera, detect the grayscale values ​​of the same coordinate position in the image between adjacent frames, and calculate the grayscale difference of each pixel between consecutive frames. Based on the grayscale difference and the coordinate index, a corresponding mapping is formed to obtain a pixel grayscale difference mapping result;

[0051] Obtain a continuous grayscale image sequence captured by the robot's forward camera. The frame rate of the image sequence is 25fps, and the size of each frame is set to 640×480. Extract the grayscale value sequence of all pixels in the image. Take an arbitrary coordinate point (x=320, y=240) as an example. Assume that the grayscale values ​​of the first and second frames corresponding to the coordinate point are 130 and 144 respectively, then the grayscale difference is 14. In actual operation, for each pair of continuous frames, perform grayscale value difference calculation on all pixels, perform pixel comparison point by point, and obtain a complete grayscale difference array ΔG, where ΔG(i, j, t) = |G( i, j, t+1)-G(i, j, t)|, where (i, j) represents the pixel position and t is the frame number. By traversing all pixel coordinate combinations in each frame image, a mapping relationship between position and grayscale difference is established. For example, the grayscale differences of 50 sample points in a certain frame image are 12, 9, 21, 5, 17, 33, 8, 14, 18, 20, etc., and a complete pixel grayscale difference mapping matrix MΔ is formed through statistics. The grayscale difference of each pixel point between frames is used as the basis for dynamic judgment. In order to facilitate the verification of the generation of actual values, a sample frame difference test group is set, see the table below:

[0052] Table 1 Grayscale difference mapping sample table

[0053] Pixel coordinate X Pixel coordinate Y Frame T Frame T+1 Gray value T Gray value T+1 Grayscale difference 320 240 1 2 130 144 14 300 220 1 2 122 126 4 280 210 1 2 134 118 16 360 200 1 2 119 135 16 310 230 1 2 112 127 15

[0054] As shown in Table 1, the grayscale values ​​of multiple coordinate points in the captured image show different degrees of change in different frames, and the grayscale difference ranges from 5 to 16, forming a sample set with obvious contrast differences. The mapping matrix data will serve as the basis for subsequent judgment of pixel change stability and obtain the pixel grayscale difference mapping result.

[0055] S112: Based on the pixel grayscale difference mapping result, perform OTSU adaptive threshold judgment on all pixel grayscale differences according to the principle of maximizing inter-class variance, filter out pixels below the threshold, determine the corresponding coordinates as static background feature points, and eliminate the pixel coordinates to obtain a coordinate set of the dynamically changing area;

[0056] Based on the pixel grayscale difference mapping results, a range analysis is performed on the grayscale difference of each pixel to identify the stable background area in the image. The OTSU threshold is set by maximizing the inter-class variance. For example, the grayscale difference sequence of the sampling points is: 4, 5, 6, 12, 16, 25, 29, 36, 39, 42, 46. After statistical frequency distribution and inter-class variance calculation, the grayscale difference threshold corresponding to the maximum inter-class variance is set to T = 18. After setting this threshold, all pixel differences are compared with this threshold. When ΔG(i, j) < 18, it is determined to be a static background. The corresponding coordinate position is deleted, and all coordinates greater than or equal to 18 are retained. For example, after processing a 640×480 image with a total of 307,200 pixels, 235,000 coordinate points are eliminated, and only 72,100 dynamically changing coordinate points are retained to form a dynamic region coordinate set S, where each coordinate corresponds to a dynamic response point in the original image. The static area is completely filtered out, and the dynamic region coordinate set is obtained.

[0057] S113: performing point-by-point statistics on the pixel difference values ​​of each coordinate in the retained area according to the dynamically changing area coordinate set, summarizing the pixel difference data and combining it with the number of coordinates to form a comparison value, thereby obtaining the number of dynamic grayscale change points;

[0058] According to the coordinate set of the dynamic change area, the grayscale differences of all retained coordinates are statistically analyzed. A point-by-point analysis method is used. The grayscale differences corresponding to the coordinate points P1 to Pn in the dynamic change area are 20, 21, 24, 30, 28, 33, 35, 29, 26, and 27, respectively. With this value sequence as input, the total statistical difference is 273, the number of points in the area is 10, and the total number of image pixels is 307,200. The difference intensity ratio is preliminarily calculated to be 0.000889. In addition, the mean position coordinates calculated according to the coordinate distribution are set to (320, 240). The Euclidean distance difference between each point position and the center point is weighted and summarized, and finally the quantitative characteristics used to reflect the grayscale response amplitude are generated. The aggregation calculation process shows that the difference trend change is concentrated in the coordinate area x = 310330, y = 235245. Further, the overall quantitative characteristics of the image dynamic response point set are extracted based on the total number of change points to obtain the number of dynamic grayscale change points.

[0059] See also Figure 3 , the specific steps of S2 are:

[0060] S211: Based on the image coordinate set corresponding to the number of dynamic grayscale change points, perform edge extraction on the original image using Canny edge detection, extract boundary segments and mark boundary pixels, count the frequency of feature points on each boundary segment and map them to boundary coordinate positions to obtain boundary feature occurrence frequency data;

[0061] Based on the image coordinate set corresponding to the number of dynamic grayscale change points, the coordinate point set of all the dynamic response areas in the image is obtained. The Canny edge detection operation is performed on the original image for the coordinates marked by the set, and the pixel points that meet the gradient double threshold conditions in the image are identified as edge candidate points. The high threshold of edge detection is set to 80 and the low threshold is set to 40. If the gradient value of a pixel point is 90, it is considered a strong edge point. If it is 45 and adjacent to a strong edge point, it is retained, otherwise it is discarded. After completing the above operation, the image boundary line segment set is formed, and each pixel on the boundary line segment is divided into unit lengths, with a unit length of 1 pixel. To divide the segment length, the boundary line segment is divided into multiple continuous segments according to the order of coordinate points. Each segment corresponds to a number of pixels. The coordinates of the dynamic points appearing in each segment are matched and counted, and the statistical frequency value is used to represent the response of the feature points within the unit length. For example, a boundary line segment is divided into 30 segments, of which the 5th segment contains 3 dynamic points, the 12th segment contains 0 dynamic points, and the 25th segment contains 2 dynamic points. By traversing all boundary segments and counting the number of dynamic coordinates within the unit length, the response frequency information of the feature points in each segment is obtained, and the mapping relationship between the boundary coordinates and the frequency value is constructed to obtain the frequency data of boundary feature occurrence.

[0062] S212: Based on the frequency data of boundary feature occurrences, the gradient direction change value of each boundary segment is calculated, the amplitude distribution is determined, a directional histogram is constructed, the frequency density corresponding to the directional distribution is counted, the directional gradient histograms on each boundary segment are summarized and averaged, and boundary directional feature density data is generated;

[0063] Based on the frequency data of boundary features, the directional gradient value is calculated according to the grayscale distribution of pixels corresponding to the boundary coordinates. The boundary directional features are represented by the direction and amplitude of the grayscale difference. The gradient angle and gradient intensity of adjacent pixels are obtained for each boundary segment. The angle interval is divided into 9 subintervals from 0 to 180 degrees, and each interval is 20 degrees wide. The gradient direction of the pixels on each boundary segment is included in the corresponding interval to form a directional distribution histogram. The boundary directional feature density of the segment is calculated based on the frequency shown in the directional histogram. For example, the directional distribution of a boundary segment is [12, 8, 15, 6, 9, 7, 3, 2, 1], then its sum is 63, and the density is the average frequency of each degree value, which is about 0.35. Then, the directional feature density value of all boundary segments is summarized and calculated. The boundary directional distribution response intensity is formed by averaging. Finally, the average directional feature of all boundary segments is extracted to obtain the boundary directional feature density data.

[0064] S213: Based on the boundary directional feature density data and in combination with the ORB feature point matching judgment standard, the difference between the directional feature value of each boundary segment and the average feature value of the overall image is judged, the deviated areas are filtered out and the remaining boundary segment data is retained. The coordinates and feature information sets of the retained boundary segments are reorganized, and a unified feature data index is established to obtain the retained boundary feature data;

[0065] According to the boundary directional feature density data, the directional feature density value of each boundary segment is compared with the density mean of all boundary segments in the image, the difference is calculated, and the abnormal boundary segments are screened out based on the difference. The boundary segments with a difference greater than 15 from the average value will be excluded. This value is used as the difference judgment threshold set by the ORB feature matching standard. The reference boundary directional density distribution standard deviation range is 10-20. The fixed threshold is 15 selected based on the experiment. For example, the directional density value of a boundary segment is 62, the overall average is 44, and the difference is 18. Therefore, it is determined that the segment needs to be eliminated, and the boundary segments within the range of ±15 near the density mean are retained. The coordinate information and feature frequency information of the boundary segments are retained, and a new feature point set is constructed. The set is matched with the original image coordinates and re-sorted. Finally, the valid data of all filtered boundary segments are integrated to generate boundary feature retention data.

[0066] Table 2 Boundary segment directional feature difference screening table

[0067]

[0068]

[0069] As shown in Table 2, based on the comparison between the difference and the set screening threshold of 15, only when the difference between the directional feature density value of the boundary segment and the overall average value is less than or equal to 15, it is judged as a valid area segment and participates in the construction of the final boundary feature set to obtain the boundary feature retention data.

[0070] See also Figure 4 , the specific steps of S3 are:

[0071] S311: Retain the image frame corresponding to the data based on the boundary features, set the ROI horizontal scanning band, extract the grayscale value distribution of all pixels in the area in each scanning band according to the image column direction, convolve each pixel in the image coordinate area to extract the edge gradient response, obtain the pixel with the maximum Sobel gradient value in each column and record the amplitude value, and obtain the scanning band gradient amplitude sequence group;

[0072] According to the boundary features, the image frame corresponding to the data is retained, the ROI horizontal scanning band is set and the area is divided according to the percentage of the image height. The sample image with a vertical size of 720 pixels is used, and the scanning band is set to 10% of the height, that is, the width of each band is 72 pixels. It slides downward from the top of the image in sequence, and each movement step is 36 pixels to form an overlapping area. Each scanning band covers all columns of the image with a column width of 1280 pixels. Grayscale reading is performed on the pixels in each scanning band and a 3×3 local window is constructed based on each pixel point. The pixel value set required by the Sobel operator is extracted, and the gradient amplitude is calculated according to the x-direction template [-101; -202; -101], and the position and amplitude value of the window center point in the image column direction are recorded. For example, the pixel grayscale value matrix of the 3×3 area corresponding to the column index 500 in the 10th row of the scanning band is [[12 3, 125, 130], [121, 124, 127], [120, 123, 126]]. The Sobel-x gradient response after convolution is (-1×123+0×125+1×130)+(-2×121+0×124+2×127)+(-1×120+0×123+1×126)=7+12+6=25. The obtained amplitude value 25 is the Sobel response amplitude of the column in the row. Continue to process all pixels along the column direction and record their maximum amplitude to form a gradient amplitude sequence in the column width direction. For example, the maximum Sobel amplitude values ​​in the range of columns 400 to 410 in the image are [12, 15, 18, 21, 20, 23, 25, 22, 19, 16, 14] respectively. Repeat the process for all scan strips, extract the corresponding position data and summarize them to obtain the scan strip gradient amplitude sequence group.

[0073] S312: Based on the scanning band gradient amplitude sequence group, perform difference calculation on adjacent column values ​​in each sequence, construct a sliding window and calculate the approximate second-order grayscale derivative of the corresponding pixel, perform Laplacian response calculation on each column, and simultaneously calculate the variance of the obtained Laplacian value. Perform mean normalization processing on the variance results of all sliding windows and sort them by column index to obtain a Laplacian variance sequence;

[0074] Based on the scanning band gradient amplitude sequence group, the Sobel amplitude values ​​on adjacent columns in each sequence are paired and the difference is calculated. A sliding window of length 3 is constructed. For example, columns 1 to 3 form the first window, columns 2 to 4 form the second window, and so on. The Laplacian approximation is calculated for each pixel inside the window, that is, the second-order derivative approximate response is calculated based on the difference between the middle column value and the left and right column values. The response value of each window position is recorded, and the variance of the response value set is calculated. The variance calculation formula is: Suppose the Laplacian response value sequence in a window is [18, 22, 24], its mean is 21.33, and the variance is ((18-21.33) 2 +(22-21.33) 2 +(24-21.33) 2 ) / 3≈6.22. After completing the single-window variance calculation, slide one column width step to the right to continue the calculation. Record all variance values ​​in column index order, and build a correspondence between the complete column index and the Laplacian variance. Perform normalization on all variance values, and set the normalization range to 0-1 to facilitate subsequent sorting and standardization judgment. Finally, arrange the image column index in ascending order to form a column-wise clarity variance distribution data structure and obtain the Laplacian variance sequence.

[0075] S313: Based on the Laplacian variance sequence, perform a difference on the variance values ​​between two adjacent columns and calculate the offset between the absolute value of all differences and the average difference. Normalize all offsets and perform a comprehensive weighting based on the Sobel amplitude change between columns. Then, sort the change rate of each column using the formula:

[0076]

[0077] Calculate the gradient change response value V g , establish the difference intensity distribution map, and obtain the gradient change amplitude distribution data, where L i represents the Laplacian variance value of the i-th column, S i Represents the Sobel amplitude value of the i-th column, M i Indicates the mean grayscale value of the scanning band in the i-th column, Represents the average value of the column grayscale mean in all scan bands, i is the column index, is the mean of all column indexes, n represents the number of columns;

[0078] According to the Laplacian variance sequence, the variance values ​​between two adjacent columns are differentially calculated and the absolute value is taken as the first-level feature change index. At the same time, the Sobel response amplitude value and the pixel grayscale mean corresponding to each column are extracted. The overall average value of the column grayscale mean and the column index average value are combined as a global reference to construct a multi-index combination difference expression. The image edge response and clarity fluctuation are analyzed column by column. For example, the sample data is set within the image column index range as follows: the Laplacian variance value of the first column is 0.15, the Sobel amplitude value is 22, and the grayscale mean is 110. The variance value of the second column is 0.24, the amplitude value is 26, and the grayscale mean is 115. The subsequent columns are: 0.20, 19, and 108 for the third column, 0.18, 17, and 112 for the fourth column, and 0.14, 16, and 106 for the fifth column. The following multiple index fusion operations are performed based on this data set: first calculate the overall grayscale mean Column index mean Then construct a mixed weighted expression for each column, combine the formula, where n = 4, and calculate each value separately: when i = 1, |L2-L1| = 0.09, The denominator is 3, and the term value is approximately 7.36;

[0079] When i=2, the difference is 0.04, The denominator is 2, and the term value is 13.25; when i=3, the difference is 0.02, The denominator is 1, and the term value is 19.15; when i=4, the difference is 0.04, The denominator is 2, the term value is 8.57, and the final sum is V g =7.36+13.25+19.15+8.57=48.33, which is the gradient change response value under the joint construction of multiple features in the current image ROI area. Combined with the distribution state of this value at the column index position and its corresponding amplitude gradient value range, an edge response trend map in the column direction is constructed to obtain the gradient change amplitude distribution data.

[0080] The gradient change response value reflects the comprehensive structural mutation intensity and edge activity of a specified area in the column direction in the image. This value integrates three types of features: image clarity fluctuation measured by Laplacian variance, edge strength expressed by the Sobel operator, and the degree of deviation of column grayscale distribution. Through unified scale weighting and position normalization, factors such as edge mutation, brightness jump, and texture concentration in the image are integrated into a quantifiable indicator. The larger the index, the more dense or drastic the edge and grayscale changes in the region are in space, and the easier it is to form stable structural feature lines or target boundaries. Therefore, it can serve as an important reference for subsequent regional saliency judgment, edge extraction, or target recognition.

[0081] The operational logic of the formula is to integrate multiple image feature dimensions to jointly measure the intensity of gradient changes, where the first term |L i+1 -L i | represents the Laplacian variance variation between adjacent columns, reflecting the sudden change in image clarity in space and serving as a direct indicator of image edge structure jump; the second term Combine the Sobel edge strength and grayscale offset of the current column, and calculate the Euclidean distance by taking the square root of the sum of squares to measure the combined effect of local edge strength and brightness distribution on the degree of mutation; the overall term is used as the difference between the numerator and the spatial position. The value after adding one is normalized as the denominator to form a spatial attenuation factor, which suppresses the influence of distant regions on the global index. The above three types of parameters are added together to form the total variation contribution, and then the weight is adjusted according to the column index distance. Finally, the regional response intensity is synthesized in the form of summation to form a multi-source fusion structural change measure.

[0082] See also Figure 5 , the specific steps of S4 are:

[0083] S411: Based on the gradient change amplitude distribution data, read the gradient amplitude value of each column one by one, and continuously slide according to the column index to determine whether it is lower than the RANSAC inlier threshold. Record the column index sequence that continuously satisfies the gradient amplitude value lower than the threshold, record the column segment length and establish the corresponding index structure to obtain the continuous low-amplitude column segment length set;

[0084] Based on the gradient change amplitude distribution data, the gradient amplitude value in the image column index direction is read column by column. The standard image area with an image resolution of 1280×720 is selected as the processing object. The image is scanned and extracted column by column. The Sobel gradient amplitude value of each column is recorded to determine whether it is lower than the set RANSAC inlier threshold. Here, the threshold is set to 1.0 pixel. Each scan band data is continuously slid. If all gradient amplitude values ​​in a certain continuous column are less than the threshold and the number of columns continuously reaches the set threshold, it is marked as a potential weak response area. The minimum continuous column number threshold is set to 5. When it is detected that 7 columns from the 350th column to the 356th column are all less than 1.0 pixel, the area is identified as a candidate weak gradient path segment. At the same time, the starting column and end column of the column segment, as well as the number of continuous columns, are recorded. After processing all scan band data, a complete low-amplitude segment length index record structure is formed, as shown below:

[0085] Table 3 Continuous low amplitude segment table

[0086] Starting column index End column index Number of consecutive columns 350 356 7 410 416 7 620 625 6

[0087] As shown in Table 3, a total of 3 continuous segments were detected, all of which met the requirements that the gradient amplitude was lower than the threshold and the continuous length met the set requirements, and the continuous low-amplitude segment length set was obtained.

[0088] S412: Based on the set of continuous low-amplitude column segment lengths, set Hough transform line segment detection parameters to determine the length of each column segment. For all column segments that meet the conditions, their start and end column coordinates and lengths are recorded one by one, and grayscale difference sequences within the range are extracted to obtain a candidate line segment grayscale difference sequence group.

[0089] According to the set of continuous low-amplitude column lengths, the minimum projection length reference value required for line segment detection in Hough transform is set, the ρ accuracy is 1 pixel, the minimum line segment projection length is set to 5 pixels, and the θ accuracy is 1 degree. It is judged whether each column segment meets the length judgment condition. Segments with a length less than 5 will be excluded. Here, the three segments of data in Table 7 are all greater than 5 columns and all meet the conditions. The original gray value sequence of the image within each column segment is further extracted, and the gray value difference is calculated as the basis for the structural change of the candidate segment. For example, the pixel gray values ​​of columns 350 to 356 are [118, 117, 119, 126, 133, 136, 138] respectively, forming a difference sequence of [-1, +2, +7, +7, +3, +2]. Then, the gray difference sequences of columns 410 to 416 and columns 620 to 625 are extracted and recorded respectively to obtain the following data:

[0090] Table 4 Candidate line segment grayscale difference data table

[0091] Column segment number Column index range Gray value sequence Grayscale difference sequence Section 1 350–356 [118，117，119，126，133，136，138] [-1，+2，+7，+7，+3，+2] Section 2 410–416 [123，121，120，127，132，133，134] [-2，-1，+7，+5，+1，+1] Section 3 620–625 [111，112，116，120，124，127] [+1，+4，+4，+4，+3]

[0092] As shown in Table 4, the grayscale change trend and difference sequence information of each segment are extracted, its response segment index is established, and the candidate segment grayscale difference sequence group is obtained.

[0093] S413: Based on the candidate line segment grayscale difference sequence group, a linkage scoring is performed according to the segment sequence length, grayscale difference change intensity, and Sobel gradient change value, using the formula:

[0094]

[0095] The operation obtains the candidate column saliency score value, performs sorting in the column segment according to the score value, determines the column index with the maximum significance as the target guiding point of the current segment, and generates the candidate guiding point position record, where ΔG j Indicates the grayscale difference of the jth column, L j is the Laplacian variance value in the continuous gradient low amplitude segment corresponding to the jth column, S j represents the Sobel amplitude value of the jth column, is the average Sobel amplitude of the current candidate segment, T k is the length of the kth candidate segment, m is the number of participating columns in the candidate segment, Z k represents the significance score of the candidate guiding point in the corresponding column segment;

[0096] According to the candidate line segment grayscale difference sequence group, the maximum jump position of the grayscale value is calculated for each sequence. At the same time, three types of feature information, Laplacian variance, Sobel amplitude, and grayscale mean offset, are introduced. Assume that the grayscale values ​​of the 350th to 356th columns in the current image segment 1 are [118, 117, 119, 126, 133, 136, 138], the grayscale difference is [-1, 2, 7, 7, 3, 2], and the maximum change Appears between columns 353 and 354, column index is 353, and its Laplacian variance is [0.15, 0.14, 0.16, 0.20, 0.23, 0.22, 0.21], Sobel amplitude is [18, 20, 22, 25, 28, 26, 24], grayscale mean is [117, 118, 120, 126, 131, 134, 136], and the overall segment mean is Use the following formula:

[0097]

[0098] Where: ΔG j is the grayscale difference, L j is the Laplacian variance value, S j is the Sobel amplitude, For average T k =7.

[0099] The specific calculation is as follows:

[0100] Item 1: |ΔG1|=1, Term value ≈ 1 + 0.0225 - 2.302 = -1.2795;

[0101] Item 2: |ΔG2| = 2, Term value ≈ 2 + 0.0196 - 1.817 = 0.2026;

[0102] Item 3: |ΔG3|=7, Term value ≈ 7 + 0.0256 - 1.140 = 5.8856;

[0103] Item 4: |ΔG4| = 7, Term value ≈ 7 + 0.0400 - 1.304 = 5.736;

[0104] Item 5: |ΔG5|=3, Term value ≈ 3 + 0.0529 - 2.167 = 0.8859;

[0105] Item 6: |ΔG6| = 2, Term value ≈ 2 + 0.0484 - 1.643 = 0.4054;

[0106] Item 7: |ΔG7| = 0, Item value ≈ 0 + 0.0441 - 0.548 = -0.5039;

[0107] sum:

[0108] -1.2795+0.2026+5.8856+5.736+0.8859+0.4054-0.5039=11.3321;

[0109] Denominator:

[0110] Final Rating:

[0111]

[0112] This score is used to measure the strength of the joint response of the grayscale fluctuations of the current column segment, and intuitively reflects the degree of aggregation of its edge features.

[0113] Table 5. Feature response data of candidate segments

[0114] Column index Grayscale difference Laplacian variance Sobel magnitude Grayscale mean deviation squared Item Value 350 -1 0.15 18 117 81 -1.2795 351 2 0.14 20 118 64 0.2026 352 7 0.16 22 120 36 5.8856 353 7 0.20 25 126 0 5.7360 354 3 0.23 28 131 25 0.8859 355 2 0.22 26 134 64 0.4054 356 0 0.21 24 136 100 -0.5039

[0115] As shown in Table 5, the current column segment score is 0.805, and the column index corresponding to the point with the maximum grayscale change is 353. It is determined to be a path guide point, and a candidate guide point position record is generated.

[0116] The candidate column significance score value represents a comprehensive score of the degree of structural mutation formed by each column in a candidate column segment of the image under the joint action of multidimensional features such as grayscale change, image clarity and edge response. This score quantifies the significance of column changes in the local area by integrating key indicators such as the grayscale difference of each column, the Laplacian variance square and the Sobel response deviation, and controls the stability of the score through segment length normalization. A higher value indicates that there are more drastic grayscale changes, clarity concentration and edge mutation phenomena in the column segment. Therefore, this score value can be used as an important indicator to measure the candidate quality of image path points and provide a quantitative basis for the subsequent extraction of significant path points.

[0117] The operational logic of the formula reflects the scoring strategy of multi-source image feature fusion. The numerator is constructed by combining the data of each column at the item level, and constructing the joint response expression of gradient change, image clarity and edge intensity offset. Specifically, |ΔG j |It is used to measure the degree of grayscale jump between adjacent columns and reflect the drastic change of local structure. is the square of the Laplacian variance, emphasizing the role of image clarity changes in scoring and improving the ability to suppress edge blur areas. It is used to describe the deviation of the Sobel gradient response from the overall average value. The square root operation is used to suppress the fluctuation caused by high-frequency abnormal responses. At the same time, the penalty trend for abnormal edge responses is reflected in the subtraction. The sum of the three items constitutes the comprehensive contribution of each column in terms of target significance. The total linear index of the overall score of the candidate segment is obtained after summing up. The denominator is composed of the column segment length and its square sum, and the The structure is normalized, where T k The square root operation is introduced to smooth the score fluctuations between segments of different lengths, avoiding false high scores in short segments due to occasional strong jumps, thus forming a candidate segment scoring system with the three-dimensional fusion of gradient strength, clarity stability and local response consistency.

[0118] See also Figure 6 , the specific steps of S5 are:

[0119] S511: Extract the pixel positions of the marked guide points in each frame based on the image coordinates in the candidate guide point position record, set a world coordinate reference template accordingly, construct a corresponding point set between the image coordinates and the world coordinates, map the positions of all image guide points in the image coordinate system to the 3D world coordinate system, and generate a camera calibration transformation coordinate set;

[0120] According to the image coordinate information in the candidate guide point position record, the two-dimensional pixel position of the key guide point marked in all image frames is read, and the corresponding physical space world coordinate point in the preset plane calibration plate template is used as a mapping reference. Coordinate mapping and pairing are performed for each group of image points and world points, and the corresponding point pairs are extracted frame by frame to solve the projection matrix. The corresponding extrinsic rotation matrix and translation vector are calculated in each frame image. In this process, the plane projection difference between the row and column coordinates of the image point and the two-dimensional plane coordinates of the actual world unit on the calibration plate are calibrated to achieve the mapping transformation of each image point to the three-dimensional coordinate. After processing multiple frames of images, all three-dimensional space coordinates under each image frame are obtained, and a point cloud structure under a unified coordinate system is established. The spatial transformation of the image guide point from the two-dimensional pixel system to the three-dimensional physical coordinate system is completed. To demonstrate the specific point position change process, the following records are listed:

[0121] Table 6 Image point calibration transformation coordinate table

[0122] Image frame number Pixel coordinates (x, y) World coordinates (X, Y, Z) F001 (523，410) (120.0，45.0，0.0) F002 (521，408) (119.5，45.2，0.0) F003 (519，407) (119.1，45.5，0.0)

[0123] As shown in Table 6, after the coordinate transformation, the key guidance points in different image frames all obtain the corresponding world coordinate positions, and the camera calibration transformation coordinate set is obtained.

[0124] S512: Based on the camera calibration transformation coordinate set, the coordinates are reordered according to the time sequence of the image frames to which the guide points belong. The three-dimensional coordinate points in adjacent time frames are extracted, and the position interval and direction change are calculated segment by segment. The spatial position difference information between each pair of coordinate points is classified into a set of path segment description data. All path segments are sequentially connected to form an ordered spatial path data structure to obtain a path trajectory set.

[0125] Based on the camera calibration transformation coordinate set, all converted three-dimensional space points are sorted by image frame number, the coordinate points between consecutive frames are paired, and their spatial position changes are calculated. The main operations include extracting the coordinate difference between the points in the previous and next frames and calculating the spatial displacement value of each segment based on the Euclidean distance calculation method. The angle between the straight line direction vectors formed by each two points is then judged, and the path deflection direction and amplitude are extracted to construct a structural data set of the path direction between segments. The starting point, end point, spacing and direction change are recorded in segments to form a set of path segment data connected in time. Subsequently, all segment information is spliced ​​to form a complete path point column. To assist in explaining the path change process, the following table is listed:

[0126] Table 7 Three-dimensional path segment data table

[0127] Segment number Starting point coordinates (X, Y, Z) End point coordinates (X, Y, Z) Spacing (unit) Path deflection angle (degrees) 1 (120.0，45.0，0.0) (119.5，45.2，0.0) 0.54 13.2 2 (119.5，45.2，0.0) (119.1，45.5，0.0) 0.50 14.8

[0128] As shown in Table 7, the current path structure already contains complete inter-segment spatial displacement and direction information, forming a standard path unit set and obtaining a path trajectory set.

[0129] S513: Based on the three-dimensional point set structure of all continuous segments in the path trajectory set, the connection relationship between the path segments is screened, and the path segments with continuous time intervals and spatial distance changes within a set range are marked as connectable segments. The inter-segment connection processing is performed in sequence, and the coordinate point sequences of all continuous segments are integrated to establish the visual robot navigation planning path;

[0130] According to the connection relationship between all segments in the path trajectory set, valid path segments are screened according to the time interval and spatial continuity requirements of the image frame numbers, and it is determined whether the segments are continuous under the preset connection threshold conditions. The threshold is set by the image frame rate and spatial displacement. Generally, the time difference between frames is set to no more than 0.3 seconds and the spatial jump between segments is set to no more than 2.0 unit lengths. All path segments that meet the conditions are sequentially merged, the end points and starting points of adjacent segments are merged, and their path numbers are adjusted and unified. After the continuous segments are spliced, a complete three-dimensional coordinate sequence structure is formed. This structure is used as the main sequence data of the navigation path for the downstream path tracking and mobile control modules to call, and the visual robot navigation planning path is established.

[0131] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A vision-based robot navigation method, characterized in that: The following steps are involved: S1: Obtain the continuous grayscale images captured by the robot's forward camera, detect the grayscale difference value of each pixel in the image at the same coordinate, mark the static background feature points and remove the corresponding coordinates, and generate the number of dynamic grayscale change points; S2: Based on the image coordinate set corresponding to the number of dynamic grayscale change points, combined with the edge detection boundary, the frequency of occurrence of boundary feature points is detected, the average feature density within the boundary is calculated, the segments below the feature matching threshold are screened out and the feature set is reconstructed to generate boundary feature retention data; S3: according to the image frame corresponding to the boundary feature retention data, set a horizontal scanning band, extract the gradient amplitude sequence, calculate the variance change rate, and generate gradient change amplitude distribution data; S4: Based on the gradient change amplitude distribution data, determine the column segments in the sequence whose continuous gradient values ​​are lower than the inlier threshold, select the column coordinates corresponding to the position with the most drastic grayscale difference change as the candidate path guide point, and generate a candidate guide point position record; S5: performing coordinate conversion according to the image coordinates in the candidate guide point position record, integrating all converted coordinate points in chronological order, constructing a path trajectory set, and obtaining a visual robot navigation planning path.

2. The vision-based robot navigation method according to claim 1, characterized in that: The number of dynamic grayscale change points includes the distribution of image active areas, pixel change intensity statistics, and initial contour information of motion boundaries. The boundary feature retention data includes local structure density distribution, edge feature stability description, and candidate navigation element screening results. The gradient change amplitude distribution data includes image texture clarity layering information, edge response intensity comparison data, and image structure mutation trend information. The candidate guidance point position records include a candidate set of navigable channel coordinates, image path guidance key point distribution, and spatial consistency verification reference values. The visual robot navigation planning path includes a time series path point set, an environment space model mapping result, and a dynamic path planning reference trajectory.

3. The vision-based robot navigation method according to claim 1, characterized in that: The specific steps for obtaining the number of dynamic grayscale change points are: S111: Obtain a continuous grayscale image sequence captured by the robot's forward camera, detect the grayscale values ​​of the same coordinate position in the image between adjacent frames, and calculate the grayscale difference of each pixel between consecutive frames. Based on the grayscale difference and the coordinate index, a corresponding mapping is formed to obtain a pixel grayscale difference mapping result; S112: Based on the pixel grayscale difference mapping result, perform OTSU adaptive threshold judgment on all pixel grayscale differences according to the principle of maximizing inter-class variance, filter out pixels below the threshold, determine the corresponding coordinates as static background feature points, and eliminate the pixel coordinates to obtain a coordinate set of the dynamically changing area; S113: performing point-by-point statistics on the pixel difference values ​​of each coordinate in the reserved area according to the dynamically changing area coordinate set, summarizing the pixel difference data and combining the coordinate quantity to form a comparison quantity, and obtaining the number of dynamic grayscale change points.

4. The vision-based robot navigation method according to claim 1, characterized in that: The specific steps for obtaining the boundary feature retention data are: S211: Based on the image coordinate set corresponding to the number of dynamic grayscale change points, perform an edge extraction operation on the original image using Canny edge detection, extract boundary segments and mark boundary pixels, count the frequency of feature points on each boundary segment and map them to boundary coordinate positions to obtain boundary feature occurrence frequency data; S212: Based on the boundary feature occurrence frequency data, calculate the gradient direction change value of each boundary segment, determine the amplitude distribution, construct a directional histogram, count the frequency density corresponding to the directional distribution, summarize the directional gradient histograms on each boundary segment and take the average to generate boundary directional feature density data; S213: Based on the boundary directional feature density data and combined with the ORB feature point matching judgment standard, the difference between the directional feature value of each boundary segment and the overall feature mean of the image is judged, the deviated area is screened out and the remaining boundary segment data is retained, the coordinates and feature information sets of the retained boundary segments are reorganized, a unified feature data index is established, and the boundary feature retention data is obtained.

5. The vision-based robot navigation method according to claim 1, characterized in that: The specific steps for obtaining the gradient change amplitude distribution data are: S311: According to the image frame corresponding to the boundary feature retention data, set the ROI horizontal scanning band, extract the grayscale value distribution of all pixels in the area in each scanning band according to the image column direction, convolve each pixel in the image coordinate area to extract the edge gradient response, obtain the pixel with the maximum Sobel gradient value in each column and record the amplitude value, and obtain the scanning band gradient amplitude sequence group; S312: Based on the scanning band gradient amplitude sequence group, performing difference calculation on adjacent column values ​​in each sequence, constructing a sliding window and calculating the approximate second-order grayscale derivative of the corresponding pixel, performing Laplacian response calculation on each column, and synchronously performing variance calculation on the obtained Laplacian values. The variance results of all sliding windows are mean-normalized and sorted by column index to obtain a Laplacian variance sequence; S313: Based on the Laplacian variance sequence, perform differential analysis on the variance values ​​between two adjacent columns, calculate the offset between the absolute value of all differences and the average difference, normalize all offsets, and perform comprehensive weighting based on the Sobel amplitude change between columns. Then, sort the change rate of each column, calculate the gradient change response value, and establish a difference intensity distribution map.

6. The vision-based robot navigation method according to claim 1, characterized in that: The specific steps for obtaining the candidate guidance point location record are: S411: Based on the gradient change amplitude distribution data, read the gradient amplitude value of each column one by one, and continuously slide according to the column index to determine whether it is lower than the RANSAC inlier threshold, record the column index sequence that continuously satisfies the gradient amplitude value lower than the threshold, record the column segment length and establish the corresponding index structure, and obtain the continuous low-amplitude column segment length set; S412: Based on the set of continuous low-amplitude column segment lengths, Hough transform line segment detection parameters are set to determine the length of each column segment. The start and end column coordinates and lengths of all column segments that meet the conditions are recorded one by one, and grayscale difference sequences within the range are extracted to obtain a candidate line segment grayscale difference sequence group. S413: Based on the candidate line segment grayscale difference sequence group, linkage scoring is performed according to the column segment sequence length, grayscale difference change intensity and Sobel gradient change value, and the candidate column significance score value is obtained by calculation. Sorting is performed within the column segment according to the score value, and the column index with the maximum significance is determined as the target guide point of the current segment, and the candidate guide point position record is generated.

7. The vision-based robot navigation method according to claim 1, characterized in that: The specific steps for obtaining the visual robot navigation planning path are: S511: Extracting the pixel positions of the marked guide points in each frame based on the image coordinates in the candidate guide point position record, and setting a world coordinate reference template accordingly, constructing a corresponding point set between the image coordinates and the world coordinates, mapping the positions of all image guide points in the image coordinate system to the three-dimensional world coordinate system, and generating a camera calibration transformation coordinate set; S512: Based on the camera calibration transformation coordinate set, reorder the image frames to which the guide points belong according to the time sequence, extract the three-dimensional coordinate points in adjacent time frames, calculate the position interval and direction change segment by segment, and classify the spatial position difference information between each pair of coordinate points into a set of path segment description data. All path segments are sequentially connected to form an ordered spatial path data structure to obtain a path trajectory set; S513: Based on the three-dimensional point set structure of all continuous segments in the path trajectory set, the connection relationship between the path segments is screened, and the path segments with continuous time intervals and spatial distance changes within the set range are marked as connectable segments. The inter-segment connection processing is performed in sequence, and the coordinate point sequences of all continuous segments are integrated to establish the visual robot navigation planning path.

Images