A blind path recognition method based on image processing

Abstract

The application relates to the technical field of image recognition, in particular to a blind path recognition method based on image processing. The method comprises the following steps: collecting a ground scene image, and performing noise suppression and illumination normalization processing to obtain a to-be-processed ground scene image; inputting the to-be-processed ground scene image into a pre-constructed convolutional neural network model to recognize blind path brick feature textures; dividing a blind path region in the to-be-processed ground scene image by using the blind path brick feature textures, and determining the direction gradient features of the blind path region; performing blind path direction consistency constraint based on the direction gradient features to collect blind path structured path segments; recognizing blind path fracture regions based on the direction gradient features; performing cross-frame target tracking according to the blind path structured path segments to output continuous blind path trajectories. The application improves the automatic recognition rate of the blind path region based on the image recognition technology, and enhances the continuous recognition capability of the blind path in a complex illumination and shielding environment.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a method for identifying tactile paving based on image processing. Background Technology

[0002] Existing technologies for tactile paving recognition and path extraction in guide robots primarily rely on manual annotation or static recognition methods based on traditional image thresholding and edge detection algorithms. While these methods can achieve basic tactile paving detection under ideal conditions of uniform lighting, clean ground, and no obstacles, their accuracy and robustness significantly decrease in complex outdoor scenarios (such as strong shadows, glare from puddles, leaf cover, dynamic pedestrian occlusion, or ground wear). With the increasing prevalence of accessible urban facilities, the diversity of tactile paving types (dotted, striped), color variations (yellow, gray, white), and material differences (rubber, concrete, ceramic tile) are becoming increasingly prominent. Traditional algorithms struggle to effectively address the interference of similar textures between tactile paving bricks and the background, lacking a comprehensive preprocessing mechanism for uneven lighting, noise pollution, and multi-scale texture details. This makes it difficult to accurately distinguish between genuine tactile paving bricks and pseudo-texture areas, leading to frequent breaks, false detections, or missed detections in path extraction results. This severely impacts the reliability of guide robots' path continuity planning, real-time obstacle avoidance decisions, and safe navigation. Summary of the Invention

[0003] Therefore, it is necessary for the present invention to provide a method for identifying tactile paving based on image processing in order to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, a method for identifying tactile paving based on image processing includes the following steps:

[0005] Step S1: Acquire ground scene images and perform noise suppression and illumination normalization processing to obtain ground scene images to be processed; input the ground scene images to be processed into a pre-built convolutional neural network model to identify the texture features of tactile paving bricks;

[0006] Step S2: Use the texture features of tactile paving bricks to segment the tactile paving area in the ground scene image to be processed, and determine the orientation gradient features of the tactile paving area; perform tactile paving orientation consistency constraints based on the orientation gradient features to collect structured path segments of the tactile paving.

[0007] Step S3: Identify the broken areas of the tactile paving based on directional gradient features; perform cross-frame target tracking based on the structured path segments of the tactile paving to output continuous tactile paving trajectories; use the continuous tactile paving trajectories to compensate for the broken areas and output the complete tactile paving path.

[0008] The beneficial effects of this invention are as follows:

[0009] (1) By performing noise suppression and illumination normalization on the ground scene image, the interference of ambient illumination changes and random noise is effectively eliminated, and the accuracy and stability of the extraction of tactile paving brick features are improved.

[0010] (2) By using a convolutional neural network model to extract and weightedly fuse the texture and edge features of the tactile paving bricks, the accurate division of the tactile paving area and the directional gradient analysis were realized, which improved the reliability of tactile paving path recognition.

[0011] (3) Based on the directional gradient features, the directional consistency constraint of the blind path is executed, and the structured path segments of the blind path are collected to effectively identify the broken area. The continuous blind path trajectory is generated by cross-frame target tracking, which significantly enhances the continuity and integrity of path recognition.

[0012] (4) By detecting and compensating for the boundary contours and gaps in the broken areas of the tactile paving, the complete reconstruction of the tactile paving path was achieved, providing stable and reliable path information for assisted navigation and intelligent travel systems. Attached Figure Description

[0013] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0014] Figure 1 This is a schematic diagram of the steps of a tactile paving identification method based on image processing according to the present invention;

[0015] Figure 2 This is a schematic diagram of the noise suppression process in this invention;

[0016] Figure 3 This is a schematic diagram illustrating the process of converting a ground scene image into a grayscale image in this invention;

[0017] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] The technical method of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention.

[0019] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0020] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0021] To achieve the above objectives, please refer to Figures 1 to 3 This invention provides a method for identifying tactile paving based on image processing, the method comprising the following steps:

[0022] Step S1: Acquire ground scene images and perform noise suppression and illumination normalization processing to obtain ground scene images to be processed; input the ground scene images to be processed into a pre-built convolutional neural network model to identify the texture features of tactile paving bricks;

[0023] In one embodiment, images of the ground scene are acquired using an industrial camera (1920×1080 resolution, 30Hz frame rate). The acquired images are converted into grayscale images, and a weighted average is performed using a 3×3 neighborhood Gaussian filter. Noise suppression is then achieved by updating each pixel value using the grayscale median. Subsequently, the image is divided into 32×32 pixel non-overlapping image blocks for illumination normalization. Linear offset is performed based on the brightness variance, and block boundary artifacts are eliminated. The normalized image is then input into a pre-constructed convolutional neural network model, which includes three convolutional layers and one feature fusion layer. The convolutional kernel size of each convolutional layer is 3×3, and the number of channels is 32, 64, and 128, respectively. Tactile paving texture and edge features are extracted through activation mapping and pooling operations. After feature fusion, a comprehensive tactile paving feature is generated, and the tactile paving feature texture mask is output by the tactile paving region determination layer.

[0024] In another embodiment, it is assumed that 10 frames of ground scene images are acquired, each frame being 1280×720 pixels in size; after grayscale conversion, a 5×5 Gaussian filter and a median filter are applied; the illumination is normalized and divided into 40×40 pixel blocks, with an average brightness offset of ±10 grayscale values; the number of channels in the convolutional layers of the CNN model are 16, 32, and 64, respectively, and the output tactile paving feature mask size is 320×180; approximately 500 local feature points of the tactile paving texture are extracted through pooling operations.

[0025] Step S2: Use the texture features of tactile paving bricks to segment the tactile paving area in the ground scene image to be processed, and determine the orientation gradient features of the tactile paving area; perform tactile paving orientation consistency constraints based on the orientation gradient features to collect structured path segments of the tactile paving.

[0026] In one embodiment, the tactile paving brick feature texture output by the CNN is converted into a binary mask, where pixels with texture values ​​higher than a threshold of 0.6 are marked as 1, and the rest as 0. A morphological closing operation of a 5×5 rectangular structuring element is performed on the binary mask, and the mask area and convex hull area are recorded. Pixels with areas in the range [500, 5000] and a convex hull ratio > 0.7 are selected and marked as tactile paving regions. Subsequently, the Sobel operator is used to calculate the horizontal and vertical gradients of the tactile paving regions, generating gradient direction angles and quantizing them into 16 discrete direction buckets. The cumulative amplitude of each bucket is calculated to determine the overall main direction. The direction features are fused into the local block histogram of each 8×8 pixel to obtain the directional gradient features of the tactile paving regions. Finally, adjacent pixels are traced along the main direction vector from the starting point of the tactile paving. The preset direction range can be set to a range value or a fixed value to generate continuous path segments. Paths with a length greater than 20 pixels are selected and collected as structured path segments for the tactile paving.

[0027] In another embodiment, it is assumed that 150 connected components of the tactile paving area are detected in the binary mask; about 30 isolated noise points are eliminated by morphological closing operation; the directional gradient is quantized into 12 directional buckets, each local block with a size of 10×10 pixels; about 50 path segments are obtained by tracing along the main direction, of which 35 are valid path segments with a length greater than the minimum threshold; the total length of the finally collected tactile paving structured path segments is about 4500 pixels.

[0028] Step S3: Identify the broken areas of the tactile paving based on directional gradient features; perform cross-frame target tracking based on the structured path segments of the tactile paving to output continuous tactile paving trajectories; use the continuous tactile paving trajectories to compensate for the broken areas and output the complete tactile paving path.

[0029] In one embodiment, the broken area is determined based on the consistency of the directional direction of the tactile paving path segments, and the broken segments with a length less than a preset threshold (such as 10 pixels) are marked as missing tactile paving areas; by cross-frame target tracking, the tactile paving path segments at corresponding positions in consecutive frames are matched, and the broken areas are compensated by linear interpolation and directional constraints to obtain a complete and continuous tactile paving trajectory output.

[0030] In another embodiment, it is assumed that there are a total of 25 broken areas in the tactile paving in 10 consecutive scene images, with each broken area having a length between 5 and 12 pixels; after cross-frame matching, 20 broken areas can be compensated, and the final output of the complete tactile paving path has a total length of about 5,000 pixels and a continuity coverage of about 92%.

[0031] It should be noted that this invention is applied to guide robots, enabling high-precision real-time identification of tactile paving and continuous path output in complex outdoor urban environments (drastic changes in lighting, pedestrian obstruction, water accumulation and fallen leaves), supporting autonomous tracking and navigation of the robot, and providing all-weather, safe and reliable barrier-free travel assistance for visually impaired people by combining voice prompts and vibration feedback.

[0032] Of particular importance, step S3, which involves identifying the tactile paving fracture region based on directional gradient features, includes:

[0033] The continuity score is calculated based on the directional gradient feature. If the continuity score is lower than the preset continuity threshold, it is marked as a potential break point. The boundary contour of the tactile paving bricks is extracted based on the potential break points.

[0034] In one embodiment, directional gradient features are first calculated based on the tactile paving image to form a continuity score map. The continuity score of each pixel reflects its consistency with the direction of its neighboring pixels. Pixels with continuity scores below a preset continuity threshold are marked as potential breakpoints. Subsequently, starting from the potential breakpoints, the boundary contour of the tactile paving bricks is extracted along the neighborhood direction of the pixels to generate a sequence of boundary points. This process can be accomplished by iteratively searching for neighboring pixels with consistent directions, ensuring that the boundary contour extends continuously along the edge of the bricks, while avoiding mismarking due to noise or local texture changes, thus providing an accurate boundary basis for subsequent tactile paving gap detection.

[0035] In another embodiment, assuming the tactile paving area is 200×200 pixels and the continuity threshold is set to 0.6, after calculating the directional gradient, a total of 50 potential break points were detected. Boundary contours were extracted starting from these potential break points, with each contour consisting of 30-45 pixels, resulting in a total of 6 main tactile paving brick boundary contours. Pixels with a continuity score below 0.4 in the boundary contours were removed through continuity score filtering, resulting in clear and continuous boundary contours, providing a quantitative basis for subsequent break area identification.

[0036] The gap area is identified by using the boundary contour of the tactile paving bricks; the curvature change value of the gap area is detected and the center coordinates, width and depth of the gap are recorded; the broken areas of the tactile paving are screened based on the center coordinates, width and depth of the gap.

[0037] In one embodiment, gap regions are identified based on the extracted boundary contours of the tactile paving bricks. For each segment on the boundary contour, the curvature change value is calculated. When a sudden change in curvature exceeds a preset curvature threshold, the segment is determined to be a gap region. Simultaneously, the center coordinates, width, and depth of the gap region are recorded to describe the spatial distribution and severity of the breakage. Subsequently, by filtering the gap center location, width, and depth information, the final tactile paving breakage area is determined, realizing a complete detection link from the boundary contour to the breakage area, providing precise positioning information for tactile paving maintenance or navigation optimization.

[0038] In another embodiment, it is assumed that the extracted boundary contours contain a total of 6 contours, each with a length of 30-45 pixels, and the curvature threshold is set to 0.5. Eight gap regions are detected, and the center coordinates of the gaps are recorded as follows: The corresponding gap width is pixels, depth is Pixels. Based on preset filtering rules (width ≥ 10 pixels or depth ≥ 2 pixels), 5 broken areas in the tactile paving were finally identified. This hypothetical numerical value demonstrates how to use abrupt curvature changes and geometric features to locate broken areas, and can be directly used for subsequent maintenance or path planning optimization.

[0039] Most importantly, in step S3, cross-frame target tracking is performed based on the structured path segments of the tactile paving, and the output of continuous tactile paving trajectories includes:

[0040] Calculate the centroid coordinates and direction vectors of the structured path segment of the tactile paving as target tracking features; predict the position of the path segment in the next frame based on the target tracking features; calculate the similarity score based on the position of the path segment in the next frame.

[0041] In one embodiment, continuously acquired tactile paving image frames are processed. The centroid position and direction vector of the path segment detected in each frame are calculated using geometric fitting or edge extraction algorithms. The position of the next frame is predicted using the centroid and direction of the path segment in the previous frame. The actual path segment is found within the predicted area by template matching or region search. The similarity between the predicted and actual path segments is calculated, including the angle between the direction vectors, the centroid position offset, and the shape overlap. A similarity score for each path segment is obtained, providing a reliable basis for subsequent cross-frame path correspondence, thereby ensuring the accuracy and real-time performance of tactile paving path continuity analysis.

[0042] In another embodiment, assuming 10 frames of tactile paving images are continuously acquired, with 5 path segments detected in each frame, the position of the next frame is predicted using the centroid and direction vector of the path segment in the previous frame. Then, the actual path segment is searched within the predicted area and the similarity score is calculated. The obtained similarities are 0.83, 0.91, 0.76, 0.88, 0.95, 0.79, 0.85, 0.90, 0.77, and 0.86, respectively. By setting a similarity threshold of 0.8, it is determined which path segments can be considered continuous. Statistics show that 42 path segments in 10 frames meet the threshold requirement. These values ​​can be used for subsequent trajectory generation, tactile paving continuity analysis, and abnormal break detection, providing a quantitative basis for path integrity assessment.

[0043] If the similarity score is higher than the preset association threshold, a cross-frame path segment correspondence is established and marked as the same trajectory; the cross-frame path segment correspondence is used to generate a continuous tactile paving center line and output the continuous tactile paving trajectory.

[0044] In one embodiment, the similarity score is compared with a preset threshold. For path segments that exceed the threshold, a matching relationship is established between adjacent frames. These matching path segments are marked as the same trajectory, and continuous center lines are generated by continuous interpolation using the centroid position and direction vector of the path segments. Continuous tactile paving trajectory data is output for navigation or assistive devices to guide tactile paving. At the same time, by combining the relative position changes between trajectory points, the stability and continuity of the path segments can be further analyzed, providing basic data for tactile paving integrity monitoring and abnormal breakage detection.

[0045] In another embodiment, assuming a preset association threshold of 0.8, using the similarity scores of 10 frames, a cross-frame correspondence relationship of 7 path segments is established in 10 consecutive frames. Each trajectory consists of multiple frame path segments. A continuous center line with a total length of about 2.1 meters is generated by linear interpolation. A point is sampled every 20 centimeters, resulting in a total of 11 trajectory points. By statistically analyzing the offset of each trajectory point and the angle between adjacent points, the continuity of the path and local offset can be analyzed, providing quantitative data on the integrity of the tactile paving trajectory for use as an input reference for tactile paving maintenance and intelligent assisted navigation systems.

[0046] Preferably, noise suppression in step S1 includes:

[0047] Step S11: Convert the ground scene image into a grayscale image; for any pixel in the grayscale image, within a neighborhood of size 3×3 or 5×5 centered on the pixel, use a Gaussian filter to perform a weighted average of the grayscale values ​​of all pixels in the neighborhood, and use the weighted average as the new grayscale value of the pixel.

[0048] In one embodiment, after reading the acquired ground scene image (resolution 1920×1080, RGB three channels), a weighted formula is used. The color image is converted to a grayscale image; then, for each pixel in the grayscale image, a 3×3 neighborhood is used as the Gaussian filter window. , The standard deviation is used to calculate a Gaussian weighted average of the gray values ​​of the nine neighboring pixels, and this result is used as the new gray value of the center pixel. After pixel processing of the entire image, a noise-suppressed grayscale image is obtained.

[0049] In another embodiment, assuming the ground scene image size is 1280×720 pixels, grayscale conversion is performed using the same weighting formula to obtain a grayscale image. Subsequently, a Gaussian filter is applied to each pixel using a 5×5 neighborhood. ; The standard deviation is used to calculate the variance of the grayscale values ​​of the entire image after processing. The average grayscale variance before noise suppression was 320, which was reduced to about 120 after processing. At the same time, a local average map of the grayscale image is generated for subsequent texture analysis.

[0050] Step S12: Sort all the new gray values ​​from smallest to largest to form a gray set; select the median value of the gray set and update the gray values ​​of the original pixels in the ground scene image.

[0051] In one embodiment, assuming the grayscale image has 921,600 pixels (1280×720), and the median grayscale value after sorting is 128; then, each pixel in the grayscale image is adjusted to a weighted average of its original pixel value and the median: ;in, This is the updated grayscale value, which is the new brightness value for each pixel; This is the original grayscale value of the current pixel in the grayscale image; The value is the middle value in the grayscale set. If the total number of pixels is odd, the grayscale value of the middle pixel is used; if it is even, the average of the grayscale values ​​of the two middle pixels is used. After this processing, the image brightness is more uniform, and the local brightness deviation is reduced by a maximum of about 35 grayscale values, which is beneficial for subsequent tactile paving texture extraction and feature recognition.

[0052] Preferably, the step S11 of converting the ground scene image into a grayscale image includes:

[0053] Step S111: Read the three-channel color values ​​of the ground scene image; calculate the weighted grayscale value using the three-channel color values, and determine the row and column coordinates of the pixels in the ground scene image;

[0054] In one embodiment, the red, green, and blue channel values ​​(R, G, B) of the acquired ground scene image (resolution 1920×1080, RGB three channels) are read pixel by pixel and then weighted using a formula. Calculate the grayscale value of each pixel; simultaneously record the row and column coordinates of that pixel in the image. This is done to facilitate subsequent processing and positioning. After processing, a complete grayscale matrix is ​​obtained, and the coordinate information corresponding to each pixel is saved.

[0055] In another embodiment, the ground scene image is assumed to be 1280×720 pixels in size, with a total of 921,600 pixels. The same weighting formula is used to calculate the grayscale values, assuming that the calculated grayscale values ​​range from 20 to 230, with an average grayscale of approximately 128. At the same time, the row and column coordinates of each pixel are recorded in a two-dimensional index table to provide a mapping relationship for subsequent storage and grayscale image generation.

[0056] Step S112: Determine the storage area using the row and column coordinates of the pixels, write the weighted grayscale value into the storage area, overwriting the original value; move to the next pixel coordinate, repeat the positioning and assignment operation until a grayscale image is generated.

[0057] In one embodiment, assuming the image size is 1280×720 pixels, the coordinates of each pixel are located sequentially using a cyclic method, and the grayscale values ​​are written to the storage area according to matrix indexing. After processing, a grayscale image is obtained, and the grayscale distribution of the entire image is statistically analyzed: the proportion of pixels with grayscale values ​​less than 50 is approximately 12%, the proportion of grayscale values ​​between 50 and 200 is approximately 78%, and the proportion of grayscale values ​​greater than 200 is approximately 10%. This grayscale distribution can be used to evaluate the brightness uniformity of the image and provide a basis for subsequent tactile paving texture analysis.

[0058] Preferably, the illumination normalization process in step S1 includes:

[0059] The ground scene image is divided into several non-overlapping image blocks of a fixed size, and each non-overlapping image block is used as a lighting statistical unit; the brightness variance is calculated using the lighting statistical unit.

[0060] In one embodiment, the acquired ground scene image (resolution 1920×1080, RGB three-channel) is divided into 64×36 non-overlapping image blocks, each image block being 30×30 pixels in size. For each image block, the RGB pixels are first converted to grayscale values. Then calculate the luminance variance of the image patch. ,in The total number of pixels in the image block. The average gray level within the image patch. The value is a grayscale value. The calculation result serves as the brightness distribution index for this illumination statistical unit, and is used for subsequent illumination normalization processing.

[0061] In another embodiment, assuming an image resolution of 1280×720 pixels, it is divided into 64×36 image blocks, each 20×20 pixels, for a total of 2304 illumination statistics units. The calculated luminance variance for each unit ranges from 5 to 45, with an average of approximately 18. This luminance variance is used to evaluate local illumination uniformity and serves as the basis for linear offset correction.

[0062] The difference between the brightness variance and the preset target brightness is calculated, and the non-overlapping image blocks are linearly offset and the normalized pixel values ​​are recorded. The non-overlapping image blocks are reassembled in their original positions using the normalized pixel values, and boundary artifacts are eliminated.

[0063] In one embodiment, the preset target brightness is set to 128 (grayscale value). For each image block, the deviation is calculated. , The average gray value of the image block is used, and then the gray values ​​of all pixels within the image block are linearly shifted. ;in, These are the pixel grayscale values ​​after brightness deviation adjustment; The difference between the average brightness of the image blocks and the preset target brightness is used to make the average brightness of each block close to the target value. All normalized pixel blocks are reassembled according to their original positions, and boundary smoothing processing (such as 5-pixel weighted average transition) is applied to eliminate inter-block artifacts, thereby obtaining a grayscale image with overall balanced illumination.

[0064] In another embodiment, assuming the maximum brightness deviation is among 2304 image blocks. minimum The average deviation is approximately 18. After performing a linear offset on each image block, the normalized grayscale value matrix is ​​recorded and reconstructed to obtain the illumination-normalized image. Statistically, the brightness values ​​in the image are concentrated in the range of 120–135, accounting for approximately 78%, while the grayscale difference in the boundary transition regions is ≤3, effectively eliminating inter-block artifacts. This processing can be used for subsequent tactile paving texture analysis or deep learning input preprocessing.

[0065] Preferably, in step S1, inputting the ground scene image to be processed into a pre-built convolutional neural network model to identify the texture features of tactile paving bricks includes:

[0066] The ground scene image to be processed is input into a pre-constructed convolutional neural network model, which includes an image feature extraction layer, a feature fusion layer, and a blind path area determination layer.

[0067] In one embodiment, a ground scene image with a resolution of 1920×1080 is acquired and input into a pre-built convolutional neural network (CNN) model. The CNN model contains three convolutional layers:

[0068]

[0069]

[0070] ;

[0071] Max pooling (2×2) is performed after each convolutional layer to extract local texture and edge features. The input image is processed through convolution and activation mapping to obtain an activation feature map, followed by pooling to extract the texture and edge features of the tactile paving bricks. The extracted features are weighted and fused through a feature fusion layer to obtain the comprehensive features of the tactile paving, which are then input into the tactile paving region determination layer to generate a tactile paving region mask. Boundary extraction and orientation labeling are also performed simultaneously, and the final output is a tactile paving brick feature texture map.

[0072] In another embodiment, assuming the input ground scene image is 1280×720 pixels and the number of channels in the convolutional layer is... The extracted activation feature map is 160×90 pixels (each feature point corresponds to 12×12 pixels in the original image). After pooling, approximately 200 local feature blocks are generated for tactile paving texture recognition. After feature fusion, a comprehensive tactile paving feature is generated, where the texture weight is assumed to be 0.6 and the edge weight is 0.4. The decision layer outputs a tactile paving region mask and marks the direction (the angle range of each mask block is 0–180°), which is used for subsequent tactile paving direction analysis and path planning.

[0073] The image feature extraction layer is used to perform convolution and activation mapping on the ground scene image to be processed to obtain an activation feature map; pooling operation is performed on the activation feature map to extract the texture and edge features of the tactile paving bricks; the feature fusion layer is used to perform weighted fusion of the tactile paving brick texture and edge features to generate comprehensive tactile paving features; the tactile paving region determination layer is used to generate a tactile paving region mask based on the comprehensive tactile paving features, and perform boundary extraction and direction marking to output the tactile paving brick feature texture.

[0074] In one embodiment, the convolutional feature map extracts local texture and edge information using 3×3 convolutional kernels, and the pooling size is 2×2 to reduce spatial resolution while retaining feature information. The feature fusion layer performs a weighted average of texture and edge features (texture 0.7, edge 0.3) to generate a comprehensive feature map. The tactile paving region determination layer uses a Sigmoid activation to output a mask, where each pixel represents the probability of a tactile paving path. The mask is obtained by binarization with a threshold of 0.5, and boundaries are extracted using Canny edge detection. Direction markers are calculated using local gradients. The final output tactile paving brick feature texture map can be used for path planning and navigation.

[0075] In another embodiment, it is assumed that the activated feature map size after pooling is approximately 160×90 pixels, with each local feature block corresponding to approximately 8×8 pixels in the original image; the texture weight is set to 0.65, and the edge weight to 0.35. The tactile paving area determination layer outputs a binary mask, in which approximately 150 tactile paving bricks are detected, with boundary lengths ranging from 30 to 80 pixels and directional angles ranging from 5° to 175°. This mask and directional information can be used for tactile paving quality assessment and robot path optimization.

[0076] Preferably, pooling is performed on the activation feature map to extract the texture and edge features of the tactile paving bricks, including:

[0077] A preset pooling window is slid across the active feature map in row and column order, and the pixels covered by each pooling window form a sub-region; pooling calculation is performed on the pixel values ​​within the sub-region, and the calculation result is used as the feature pixel value;

[0078] In one embodiment, the activation feature map size is 128×128, and the pooling window size is 4×4 with a stride of 4. Max pooling is performed on the pixels within each pooling window to obtain a pooled feature map size of 32×32. Each feature pixel value represents the local maximum activation intensity of the sub-region, thereby reducing spatial redundancy and preserving significant texture information.

[0079] In another embodiment, assume the activation feature map size is 100×100, the pooling window is 5×5, and the stride is 5. After pooling, a 20×20 pooled feature map is obtained, with a total of 400 feature pixels. The maximum value in each sub-region is recorded as the feature pixel value, and the minimum value is recorded as a contrast index for subsequent texture intensity and edge analysis.

[0080] A sliding pooling window is used to traverse the activation feature map to generate a pooled feature map; the gradient of the activation value in the pooled feature map is calculated and used as a texture intensity index; the texture of the tactile paving bricks is extracted based on the texture intensity index.

[0081] In one embodiment, the horizontal gradient is calculated for each feature pixel of the pooled feature map. and vertical gradient Through formula The texture intensity index is obtained. The location where the texture intensity index is greater than a preset threshold (e.g., 0.3) is identified as the texture area of ​​the tactile paving brick, which is used for subsequent tactile paving identification and path planning.

[0082] In another embodiment, assuming the pooled feature map is 20×20, the gradient calculation yields texture intensity values ​​ranging from 0 to 1. Setting the threshold to 0.35, approximately 120 feature pixels are marked as texture points for tactile paving, among which high-intensity pixels... Approximately 30, which can be used to assist in edge and orientation analysis.

[0083] Edge response values ​​are calculated using texture intensity indices, and potential edge pixels are labeled. Connectivity analysis is then performed based on these potential edge pixels to extract edge features.

[0084] In one embodiment, the edge response value of each texture point is calculated using the Sobel operator, and pixels with an edge response value greater than 0.25 are marked as potential edges. Subsequently, 8-neighborhood connectivity analysis is used to combine adjacent edge pixels into edge segments, extracting complete tactile paving brick boundary features.

[0085] In another embodiment, assuming there are 150 texture intensity points in the pooled feature map, approximately 60 potential edge pixels are marked after edge response calculation. Using connectivity analysis, these pixels are formed into 15 continuous edge segments, each with a length ranging from 3 to 12 pixels, and the orientation angle of each edge segment (assuming a range of 5°–175°) is calculated for tactile paving direction determination and brick localization.

[0086] Preferably, step S2, which involves using the texture features of tactile paving bricks to delineate the tactile paving area in the ground scene image to be processed, includes:

[0087] The tactile paving brick feature texture is converted into a binary mask, where pixels with tactile paving brick feature texture above a preset texture threshold are marked as 1, and the remaining pixels are marked as 0; morphological closing operation is performed on the binary mask, and the mask features are recorded.

[0088] In one embodiment, the texture features of the tactile paving bricks are converted into a binary mask, where pixels with texture features higher than a preset texture threshold of 0.35 are marked as 1, and the remaining pixels are marked as 0. Morphological closing operations are performed on the binary mask, including a 3×3 pixel dilation operation followed by an erosion operation of the same size, to eliminate small holes and connect broken areas. After processing, the boundary pixel coordinates, center point coordinates, and region number of each connected region are recorded, providing a basis for subsequent tactile paving region selection. During the closing operation, regions with a pixel spacing of less than 5 pixels between connected regions are automatically connected to ensure region continuity. This processing effectively removes noisy pixels while preserving the continuous texture features of the tactile paving bricks, making subsequent area and shape calculations more accurate.

[0089] In another embodiment, assuming the binary mask image size is 256×256 pixels and the threshold is set to 0.4, the closing operation generates approximately 1500 foreground pixels, forming 18 connected regions. The number of boundary pixels in each region ranges from 60 to 220. By recording the center point coordinates and region number of each region, the pixel area and convex hull area can be directly calculated subsequently. The connectivity and integrity of each region are guaranteed after the closing operation, ensuring that even if some regions have gaps or texture breaks, continuous regions can be formed for filtering and labeling. This method ensures that the features of the tactile paving area are accurately represented, while facilitating subsequent calculations and judgments.

[0090] Pixel area and convex hull area are calculated based on mask features; pixels in the ground scene image to be processed with a pixel area in the range [500, 5000] and a convex hull area ratio greater than 0.7 are selected and marked as tactile paving areas.

[0091] In one embodiment, based on the connected component mask features, the pixel area of ​​each region is calculated, i.e., the total number of pixels in the connected component, and the total number of pixels in the minimum convex hull coverage area is calculated to obtain the convex hull area. The convex hull area ratio is further calculated as pixel area / convex hull area. Regions that satisfy the pixel area range of [500, 5000] and have a convex hull area ratio greater than 0.7 are selected and marked as tactile paving regions. This method can exclude regions that are too small or too sparse in shape, while ensuring that continuous areas of tactile paving bricks are completely identified. The marked tactile paving regions include boundary coordinates, center point coordinates, and region number, providing input for subsequent boundary extraction and direction marking, ensuring that the entire tactile paving detection process can be executed accurately.

[0092] In another embodiment, assuming the obtained 18 connected regions have a pixel area range of [300, 6000] pixels, after filtering, 12 tactile paving regions are obtained with pixel areas of 520, 870, 1100, 1500, 2100, 2500, 3100, 3700, 4200, 4800, 5000, and 4950 pixels, respectively, corresponding to convex hull area ratios of 0.72, 0.75, 0.78, 0.81, 0.73, 0.77, 0.80, 0.85, 0.88, 0.74, 0.79, and 0.82. Each selected region completely covers the tactile paving texture, and the center point coordinates are used for further direction marking and boundary extraction. This numerical assumption can be directly used for experimental verification and algorithm performance evaluation.

[0093] Preferably, determining the directional gradient characteristics of the tactile paving area in step S2 includes:

[0094] Calculate the horizontal and vertical gradients of the tactile paving area, and generate gradient direction angles using the Sobel operator; quantize the gradient direction angles into 8-16 preset discrete direction buckets, and calculate the cumulative gradient magnitude of each discrete direction.

[0095] In one embodiment, for the identified tactile paving area, the horizontal and vertical gradients of each pixel are first calculated, and the Sobel operator is used to apply the gradients along the grayscale image. shaft and Differentiating along the axis yields and And calculate the gradient magnitude of each pixel. and gradient direction Then the gradient direction The gradient is quantized into eight discrete directional buckets, and the cumulative sum of gradient magnitudes within each bucket is calculated to describe the gradient energy distribution within that tactile paving area. This process effectively reflects the main directional characteristics of the tactile paving texture and provides fundamental information for subsequent regional directional analysis. The entire calculation process normalizes pixel values ​​to reduce the impact of illumination variations on gradient calculation while preserving texture structure features.

[0096] In another embodiment, assuming the tactile paving area is 128×128 pixels in size and the Sobel operator kernel size is 3×3, the gradient directions are quantized into 16 discrete buckets. The cumulative gradient magnitude is calculated for each bucket, resulting in 16 gradient energy values: [120, 85, 60, 45, 30, 20, 15, 10, 5, 3, 2, 2, 1, 1, 0, 0]. This cumulative statistical analysis determines the intensity distribution of the texture in a specific direction. The maximum gradient magnitude occurs in the first and second buckets, indicating that the tactile paving texture generally tends to be horizontally aligned. This assumed value provides a quantifiable reference for subsequent directional feature extraction and region comparison.

[0097] The main direction is selected based on the cumulative gradient magnitude as the overall directional feature of the tactile paving area; the tactile paving area is divided into blocks, and a local directional gradient histogram is calculated for each block; the overall directional features are fused into the local directional gradient histograms to determine the directional gradient features of the tactile paving area.

[0098] In one embodiment, based on the cumulative gradient magnitude, the direction corresponding to the largest magnitude is selected as the overall directional feature of the tactile paving area. Then, the tactile paving area is divided into several equally sized blocks (e.g., 8×8 pixels). A local directional gradient histogram is calculated within each block, and the cumulative gradient magnitude of each directional bucket is statistically analyzed. The overall directional feature is then fused with the block's local directional gradient histogram to adjust the local directional weights, making the local histogram more closely match the overall texture direction. Finally, each block obtains a fused directional gradient feature, which is used to describe the directional consistency and local texture distribution of the tactile paving area, providing reliable feature information for tactile paving detection, recognition, and direction correction.

[0099] In another embodiment, it is assumed that the 128×128 pixel tactile paving area is divided into 16×16 blocks, each block being 8×8 pixels. The overall directional feature is the second directional bucket (gradient magnitude 85). In the local directional gradient histograms calculated for each block, the maximum gradient in some blocks appears in the first and third directional buckets. By fusing the overall directional features, the magnitude corresponding to the second directional bucket for each block is weighted to 0.7, while the magnitudes of other directions remain unchanged. After fusion, the local directional gradient histogram of each block shows a gradient concentration trend, which is beneficial for subsequent directional feature analysis and identification of the tactile paving area. It is assumed that the final local directional magnitude matrix of the 16×16 blocks is used for visualization and direction determination.

[0100] like Figure 3 As shown in the figure, this diagram illustrates the distribution of cumulative gradient amplitudes across 16 discrete directional blocks in the tactile paving area. The figure clearly shows a highly concentrated attenuation pattern in the gradient energy. The first directional block has the highest cumulative gradient amplitude of 120, and the second directional block has 85. These two dominant directions account for the vast majority of the gradient energy, indicating that the tactile paving texture has a clear dominant directional characteristic, generally tending towards a horizontal arrangement. As the directional block number increases, the gradient amplitude shows a rapid attenuation trend: the first five directional blocks (120, 85, 60, 45, 30) concentrate approximately 90% of the gradient energy, while the amplitude of the latter 11 directional blocks gradually decreases to near zero. This distribution characteristic fully reflects the directional consistency and structural regularity of the tactile paving texture, providing a reliable quantitative basis for subsequent dominant direction selection, local directional gradient histogram fusion, and directional feature analysis of the tactile paving area. Through this gradient direction statistics, the laying direction of the tactile paving can be effectively identified, providing accurate feature information for direction correction and regional comparison.

[0101] Preferably, in step S2, performing tactile directional consistency constraints based on directional gradient features to collect structured path segments of the tactile directional path includes:

[0102] The directional gradient features are converted into a vector field, where each pixel is associated with a direction vector; the principal direction vector of the vector field is determined; starting from the starting point of the tactile paving area, adjacent pixels are traced along the principal direction vector, the direction deviation is calculated, and the direction deviation is checked.

[0103] In one embodiment, the obtained directional gradient features of the tactile paving area are converted into a vector field, with each pixel associated with a directional vector. The directional magnitude and directional angle of each pixel are calculated through gradient calculation. Next, the directional distribution of the entire tactile paving area is statistically analyzed, and the direction with the largest cumulative directional magnitude is determined as the main directional vector. Starting from the beginning endpoint of the tactile paving area, adjacent pixels are tracked along the main directional vector. At each step, the directional deviation between the directional vector of the candidate adjacent pixel and the main directional vector is calculated, and the directional deviation is compared with a preset directional range. The preset directional range includes a first directional threshold and a second directional threshold, where the second directional threshold is greater than the first directional threshold.

[0104] If the directional deviation is within the preset directional range, then the adjacent pixels are included in the current path segment and tracking continues;

[0105] If the directional deviation is less than the first directional threshold, and the candidate adjacent pixel is located within the blind path area and has not been visited, then it is included in the current path segment and marked as visited, and the search and judgment are repeated recursively.

[0106] If the directional deviation exceeds the second directional threshold, the current path segment is terminated and a new path segment is started; the current path is filtered by length, and reasonable path segments with a length greater than the preset minimum length are retained; the structured path segments of the tactile paving are selected based on the reasonable path segments.

[0107] Based on the above steps, basic information is provided for subsequent path segmentation, thereby forming a continuous direction vector tracking framework, laying the foundation for tactile paving path recognition.

[0108] In another embodiment, assuming the tactile paving area is 128×128 pixels, each pixel's direction is quantized into 16 discrete directions, and a direction vector field is calculated. The average direction angle of the main direction vector is 45°, and the second direction threshold is set to 12°. Pixels are tracked along the main direction from the top left corner, and the lengths of the continuously tracked paths are statistically represented as [15, 22, 18, 30, 25, 12, 28, 10, 14, 20] pixels. When the direction deviation exceeds the second direction threshold, i.e., the direction deviation of the 8th path is 14°, the current path is stopped and a new path segment is started. This assumed value is used to demonstrate how to generate an initial tactile paving path through direction vector tracking, providing a quantifiable basis for subsequent path segment selection.

[0109] In another embodiment, based on the path segments, the path segments are further filtered for consistency in length and direction. First, the average direction of the pixel direction sequence of each path segment is calculated. If the deviation between the average direction and the main direction vector is less than a first direction threshold, it is determined to be a path with consistent direction. Then, the length of the path segment is calculated, and path segments with a length less than a preset minimum length (e.g., 20 pixels) are discarded, retaining reasonable path segments. Further, the spacing between adjacent reasonable path segments is analyzed. When the pixel spacing between adjacent path segments is less than a preset pixel gap (e.g., 5 pixels) and the direction is consistent, the path segments are merged to form a continuous tactile paving structured path segment. This method can effectively retain long, continuous, and consistent tactile paving paths, eliminate noise or short fragmented paths, and generate tactile paving structured path features that are easy to identify and navigate later.

[0110] In the above embodiments, if the directional deviation of a candidate adjacent pixel is less than a first directional threshold, and the candidate adjacent pixel is located within the tactile paving area and has not been visited, it is included in the current path segment, and the pixel is marked as visited in the access marker map. Subsequently, using the end pixel of the current path segment as a new starting point, the search and directional deviation judgment are recursively repeated to gradually extend the path segment until it can no longer be extended. This method can ensure that the path grows continuously along the main direction of the tactile paving and avoid repeated visits or deviations from the tactile paving area, thereby generating complete, continuous, and directionally consistent path segments, providing basic data support for the structured analysis and navigation of tactile paving.

[0111] In another embodiment, it is assumed that 20 initial path segments are generated, with lengths ranging from [8, 12, 25, 30, 15, 10, 22, 28, 18, 16, 14, 20, 26, 9, 12, 19, 23, 17, 11, 21] pixels. A preset minimum length threshold of 15 pixels, a directional deviation of 10°, and a pixel gap threshold of 5 pixels are set. After filtering and merging, 12 continuous path segments are ultimately retained and merged, with the longest path segment being 30 pixels and the shortest being 16 pixels. These assumed values ​​illustrate how, in practical applications, length, direction, and gap constraints can be used to filter reasonable paths, thereby achieving the extraction and organization of structured paths in tactile paving areas.

[0112] In another embodiment, it is assumed that 10 initial path segments are obtained, each with a length of [12, 15, 9, 20, 14, 10, 18, 11, 16, 13] pixels. All unvisited pixels are recursively expanded, with each path segment being extended recursively in [3, 5, 2, 6, 4, 3, 5, 2, 4, 3] steps. The final generated continuous path segments have lengths of [15, 20, 12, 25, 18, 13, 22, 13, 20, 16] pixels, and duplicate or redundant paths are avoided through access markers. This hypothetical numerical demonstration illustrates the specific process of generating reasonable tactile paving path segments by combining recursive search with direction deviation judgment, providing a quantitative example for subsequent path fusion and structured processing.

[0113] Preferably, tracing adjacent pixels along the main direction vector and calculating the direction deviation includes:

[0114] Read its main direction vector and determine the unit forward direction; search for candidate neighboring pixels in the forward or reverse direction of the unit forward direction with a fixed step size; obtain the direction vector of the candidate neighboring pixel and calculate the direction deviation between the direction vector of the candidate neighboring pixel and the main direction vector.

[0115] In one embodiment, the main direction vector of the tactile paving area is read, and the unit forward direction is determined accordingly. Candidate neighboring pixels are searched along the unit forward direction in either the forward or reverse direction with a preset fixed step size (e.g., 1 pixel). The direction vector of each candidate neighboring pixel is obtained, and the angle between it and the main direction vector is calculated to obtain the direction deviation. The direction deviation is used to measure the consistency between the candidate neighboring pixels and the forward direction, and serves as a basis for judging path continuity.

[0116] In another embodiment, assuming the tactile paving area is 128×128 pixels, the initial direction of the main direction vector is... The fixed step size is 1 pixel. When searching in the forward or reverse direction along the unit forward movement, the statistical deviation angle between candidate adjacent pixels is calculated as follows: These candidate adjacent pixels are all located within the tactile paving area, initially forming 10 path segments. Based on a preset directional range, if the directional deviation is less than the first directional threshold within the preset directional range, such as when set to... Pixels with a directional deviation exceeding a second directional threshold are included in the current path segment, while pixels with a directional deviation exceeding a second directional threshold are used to initiate a new path segment. This hypothesis demonstrates how candidate paths can be constructed along the main direction at the pixel level, providing a quantitative basis for subsequent path selection.

[0117] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0118] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for identifying tactile paving based on image processing, characterized in that, Includes the following steps: Step S1: Acquire ground scene images and perform noise suppression and illumination normalization processing to obtain ground scene images to be processed; input the ground scene images to be processed into a pre-built convolutional neural network model to identify the texture features of tactile paving bricks; Step S2: Use the texture features of tactile paving bricks to segment the tactile paving area in the ground scene image to be processed, and determine the directional gradient features of the tactile paving area; Performing directional consistency constraints on tactile paving based on directional gradient features to collect structured path segments of tactile paving includes: converting directional gradient features into a vector field, where each pixel is associated with a direction vector; determining the principal direction vector of the vector field; tracing adjacent pixels along the principal direction vector from the starting point of the tactile paving area, calculating the direction deviation, and checking the direction deviation. If the directional deviation is within a preset directional range, the adjacent pixels are included in the current path segment and tracking continues. The preset directional range includes a first directional threshold and a second directional threshold, and the second directional threshold is greater than the first directional threshold. If the directional deviation is less than the first directional threshold, and the candidate adjacent pixel is located in the blind path area and has not been visited, then it is included in the current path segment and marked as visited, and the search and judgment are repeated recursively. If the directional deviation exceeds the second directional threshold, the current path segment is terminated and a new path segment is started; the current path is filtered by length, and reasonable path segments with a length greater than the preset minimum length are retained; the structured path segments of the tactile paving are selected based on the reasonable path segments; Step S3: Identify the broken areas of the tactile paving based on directional gradient features; perform cross-frame target tracking based on the structured path segments of the tactile paving to output continuous tactile paving trajectories; use the continuous tactile paving trajectories to compensate for the broken areas and output the complete tactile paving path.

2. The tactile paving identification method based on image processing according to claim 1, characterized in that, Noise suppression in step S1 includes: Step S11: Convert the ground scene image into a grayscale image; for any pixel in the grayscale image, within a neighborhood of size 3×3 or 5×5 centered on the pixel, use a Gaussian filter to perform a weighted average of the grayscale values ​​of all pixels in the neighborhood, and use the weighted average as the new grayscale value of the pixel. Step S12: Sort all the new gray values ​​from smallest to largest to form a gray set; select the median value of the gray set and update the gray values ​​of the original pixels in the ground scene image.

3. The tactile paving identification method based on image processing according to claim 2, characterized in that, The step S11 of converting the ground scene image into a grayscale image includes: Step S111: Read the three-channel color values ​​of the ground scene image; calculate the weighted grayscale value using the three-channel color values, and determine the row and column coordinates of the pixels in the ground scene image; Step S112: Determine the storage area using the row and column coordinates of the pixels, write the weighted grayscale value into the storage area, overwriting the original value; move to the next pixel coordinate, repeat the positioning and assignment operation until a grayscale image is generated.

4. The tactile paving identification method based on image processing according to claim 1, characterized in that, The illumination normalization process in step S1 includes: The ground scene image is divided into several non-overlapping image blocks of a fixed size, and each non-overlapping image block is used as a lighting statistical unit; the brightness variance is calculated using the lighting statistical unit. The difference between the brightness variance and the preset target brightness is calculated, and the non-overlapping image blocks are linearly offset and the normalized pixel values ​​are recorded. The non-overlapping image blocks are reassembled in their original positions using the normalized pixel values, and boundary artifacts are eliminated.

5. The tactile paving identification method based on image processing according to claim 1, characterized in that, Step S1 involves inputting the ground scene image to be processed into a pre-built convolutional neural network model to identify the texture features of tactile paving bricks, including: The ground scene image to be processed is input into a pre-constructed convolutional neural network model, which includes an image feature extraction layer, a feature fusion layer, and a blind path area determination layer. The image feature extraction layer is used to perform convolution and activation mapping on the ground scene image to be processed to obtain an activation feature map; pooling operation is performed on the activation feature map to extract the texture and edge features of the tactile paving bricks; the feature fusion layer is used to perform weighted fusion of the tactile paving brick texture and edge features to generate comprehensive tactile paving features; the tactile paving region determination layer is used to generate a tactile paving region mask based on the comprehensive tactile paving features, and perform boundary extraction and direction marking to output the tactile paving brick feature texture.

6. The tactile paving identification method based on image processing according to claim 5, characterized in that, Pooling operations are performed on the activation feature maps to extract the texture and edge features of the tactile paving bricks, including: A preset pooling window is slid across the active feature map in row and column order, and the pixels covered by each pooling window form a sub-region; pooling calculation is performed on the pixel values ​​within the sub-region, and the calculation result is used as the feature pixel value; A sliding pooling window is used to traverse the activation feature map to generate a pooled feature map; the gradient of the activation value in the pooled feature map is calculated and used as a texture intensity index; the texture of the tactile paving bricks is extracted based on the texture intensity index. Edge response values ​​are calculated using texture intensity indices, and potential edge pixels are labeled. Connectivity analysis is then performed based on these potential edge pixels to extract edge features.

7. The tactile paving identification method based on image processing according to claim 1, characterized in that, Step S2, which involves using the texture features of tactile paving bricks to segment the tactile paving area in the ground scene image to be processed, includes: The tactile paving brick feature texture is converted into a binary mask, where pixels with tactile paving brick feature texture above a preset texture threshold are marked as 1, and the remaining pixels are marked as 0; morphological closing operation is performed on the binary mask, and the mask features are recorded. Pixel area and convex hull area are calculated based on mask features; pixels in the ground scene image to be processed with a pixel area in the range [500, 5000] and a convex hull area ratio greater than 0.7 are selected and marked as tactile paving areas.

8. The tactile paving identification method based on image processing according to claim 1, characterized in that, Step S2, determining the directional gradient characteristics of the tactile paving area, includes: Calculate the horizontal and vertical gradients of the tactile paving area, and generate gradient direction angles using the Sobel operator; quantize the gradient direction angles into 8-16 preset discrete direction buckets, and calculate the cumulative gradient magnitude of each discrete direction. The main direction is selected based on the cumulative gradient magnitude as the overall directional feature of the tactile paving area; the tactile paving area is divided into blocks, and a local directional gradient histogram is calculated for each block; the overall directional features are fused into the local directional gradient histograms to determine the directional gradient features of the tactile paving area.

9. The tactile paving identification method based on image processing according to claim 1, characterized in that, Tracing adjacent pixels along the main direction vector, the calculation of directional deviation includes: Read its main direction vector and determine the unit forward direction; search for candidate neighboring pixels in the forward or reverse direction of the unit forward direction with a fixed step size; obtain the direction vector of the candidate neighboring pixel and calculate the direction deviation between the direction vector of the candidate neighboring pixel and the main direction vector.

Images