Waterproof board laying and hanging robot path identification method based on machine vision

Abstract

The invention provides a waterproof board laying and hanging robot path recognition method based on machine vision, and relates to the technical field of image recognition. According to the method, an industrial camera carried by a waterproof board laying and hanging robot collects a waterproof board laying and hanging area image in a tunnel, perspective transformation is performed by combining internal and external parameters of the camera, and a laying operation image under an orthographic view angle is obtained; performing Gaussian blur, contrast enhancement, brightness normalization and other preprocessing operations on the laying operation image; after the laying operation image is preprocessed, a directional entropy gradient field is constructed based on local information entropy, a disturbance response is fused to generate a structure display function, and a structure feature map of the laying operation image is extracted; a path point sequence for navigation control of the laying and hanging robot is obtained through an image skeleton extraction and path point equal-interval sampling method, noise reduction processing is conducted on the path point sequence through sliding window mean filtering, smooth fitting is conducted on the path sequence through B-spline curve fitting, and a robot motion path is obtained.

Description

Technical Field

[0001] The present invention belongs to the technical field of image recognition, and in particular relates to a path recognition method for a waterproof board laying robot based on machine vision. Background Art

[0002] Currently, the installation and hanging of waterproof sheets during tunnel structure construction mostly relies on manual labor, which presents problems such as low efficiency, high repetitive labor intensity, and high safety risks. With the growing demand for automation in tunnel projects, waterproof sheet installation robots have received widespread attention as important equipment for improving construction efficiency and reducing labor costs. However, for robots to complete precise installation and hanging operations in complex tunnel environments, they require the support of high-precision path recognition and navigation control technologies.

[0003] Traditional robot path recognition methods mostly rely on manually setting reference trajectories or using hardware methods such as lidar and inertial navigation to construct paths. However, in tunnel paving scenarios, they are restricted by construction conditions, dust obstruction, and lighting changes, which can easily cause path drift and misjudgment. At the same time, existing vision-based path recognition methods often have difficulty accurately extracting structural features from complex backgrounds, and have technical bottlenecks such as large image noise interference and low path positioning accuracy. They cannot meet the requirements of paving robots for path consistency, feasibility, and robustness.

[0004] To address the above problems, a path recognition method for a waterproof board laying robot based on machine vision is proposed. In this method, the industrial camera carried by the waterproof board laying robot is used to collect images of the waterproof board laying area in the tunnel. The camera's internal and external parameters are combined to perform perspective transformation to obtain the laying operation image under the orthographic perspective. The laying operation image is preprocessed by Gaussian blurring, contrast enhancement, and brightness normalization. After the preprocessing of the laying operation image is completed, a directional entropy gradient field is constructed based on the local information entropy, and the disturbance response is integrated to generate a structure visualization function to extract the structural feature map of the laying operation image. The path point sequence for the navigation control of the laying robot is obtained by image skeleton extraction and equidistant sampling of path points. The path point sequence is denoised by sliding window mean filtering, and the path sequence is smoothed by B-spline curve fitting to form a continuous and usable robot motion path. Summary of the Invention

[0005] The present invention provides a path recognition method for a waterproof board laying robot based on machine vision. The method comprises the following steps: an industrial camera carried by the waterproof board laying robot is used to collect images of a waterproof board laying area in a tunnel; a perspective transformation is performed in combination with internal and external parameters of the camera to obtain a laying operation image under an orthographic perspective; Gaussian blurring, contrast enhancement, brightness normalization and other preprocessing operations are performed on the laying operation image; after the preprocessing of the laying operation image is completed, a directional entropy gradient field is constructed based on local information entropy, a structural visualization function is generated by fusing disturbance responses, and a structural feature map of the laying operation image is extracted; a path point sequence for navigation control of the laying robot is obtained by image skeleton extraction and a method of equally spaced path point sampling; a sliding window mean filter is used to perform noise reduction on the path point sequence; and a B-spline curve fitting is used to smooth the path sequence to obtain a motion path of the robot.

[0006] To achieve the above-mentioned purpose, the technical solution adopted by the present invention is: a path recognition method for a waterproof board laying robot based on machine vision, and the specific steps include:

[0007] S1. The industrial camera carried by the waterproof board laying robot collects images of the waterproof board laying area in the tunnel. Perspective transformation is performed based on the internal and external parameters of the camera to obtain the laying operation image under the orthographic perspective.

[0008] S2. Performing Gaussian blur, contrast enhancement, and brightness normalization preprocessing operations on the paving operation image.

[0009] S3. Use the image entropy flow offset response method to extract features from the pre-processed paving operation image, including constructing the local information entropy of the paving operation image, constructing a directional entropy gradient field based on the local information entropy, and constructing a structure visualization function based on the disturbance response and the entropy gradient to obtain a structure feature map.

[0010] S4. Using the image skeleton extraction and path point equidistant sampling method, perform path recognition and path point generation operations on the structural feature graph to obtain a path point sequence for navigation control of the paving robot.

[0011] S5. Use sliding window mean filtering to perform noise reduction processing on the path point sequence, and perform smooth fitting on the path point sequence through B-spline curve fitting to obtain the robot motion path.

[0012] Preferably, in S2, the paving operation image is preprocessed, and the specific method is: Performing Gaussian blur processing on the paving operation image to obtain a blurred paving operation image, calculating a high-frequency residual map between the paving operation image and the blurred paving operation image, and fusing the high-frequency residual map with the paving operation image to obtain a high-frequency enhanced paving operation image; The high-frequency enhanced paving operation image is divided into multiple sub-regions of equal size. The brightness mean and structural gradient variance of each sub-region are calculated respectively. According to the brightness mean of each sub-region, the brightness normalization operation is performed on each sub-region. According to the structural gradient variance of each sub-region, the contrast adjustment operation is performed on each sub-region. The processed pixels of each sub-region are spliced ​​into a complete image to obtain the preprocessed paving operation image.

[0013] Preferably, in S2, the paving operation image is preprocessed, and the noise of the paving operation image can be effectively suppressed by the Gaussian blur operation, while retaining the key structural information. By fusing the high-frequency residual map with the paving operation image, the edge and detail features of the paving operation image can be significantly enhanced; the normalization processing based on the brightness mean can effectively eliminate the influence of uneven lighting and improve the overall consistency of the paving operation image.

[0014] Preferably, in S3, the local information entropy of the paving operation image is constructed, and the specific method is: For the pre-processed paving operation image I pre For each pixel position (x, y) in (x, y), within the neighborhood N(x, y) centered at (x, y), the distribution probability of the grayscale values ​​of all pixels in the neighborhood is counted to obtain the local grayscale probability distribution P (x，y) (v), the specific calculation process is: Where v is the pixel grayscale value, v∈[0,255], δ(·) is the indicator function, which counts the number of occurrences of each pixel grayscale value; the local information entropy E(x, y) of the pixel position is calculated based on the local grayscale probability distribution. The specific calculation process is: Where ∈ is a very small integer.

[0015] Preferably, in S3, a directional entropy gradient field is constructed based on the local information entropy, and the specific method is: Define the direction set Θ for detecting direction changes, Θ = {0°, 45°, 90°, 135°, 180°, 225°, 270°, 315°}, and for each direction θ, θ∈Θ, define the unit pixel offset (Δx θ , Δy θ ), for each pixel position (x, y), calculate the information entropy difference F of (x, y) in direction θ θ (x, y), the specific calculation process is: F θ (x, y) = E(x, y) - E(x + Δx θ , y+Δyθ ); The maximum value of the information entropy difference in all directions is taken to obtain the directional entropy gradient response intensity S(x, y) of the pixel. The S(x, y) of each pixel in the paving operation image constitutes a directional entropy gradient field.

[0016] Preferably, in S3, a structure visualization function is constructed based on the disturbance response and the entropy gradient to obtain a structure feature map, and the specific method is as follows: For each pixel, the local grayscale probability distribution P (x，y) (v) Using Gaussian distribution, construct a normal distribution with the neighborhood grayscale mean as the center to obtain the local Gaussian probability distribution Q (x，y) (v) P (x，y) (v) with Q (x，y) (v) Perform linear mixing to obtain the perturbation probability distribution P′ (x，y) (v), based on P′ (x，y) (v) recalculating the local information entropy of the pixel area to obtain the perturbed local information entropy E′(x, y), and performing a difference calculation between the original local information entropy E(x, y) and the perturbed local information entropy to obtain the perturbation response value ΔE(x, y); The directional entropy gradient response intensity S(x, y) of each pixel position is multiplied by the disturbance response value ΔE(x, y) to obtain the structural response map M(x, y). The maximum and minimum normalization operation is performed on M(x, y) to obtain the normalized structural response map M′(x, y). The normalized structural response map M′(x, y) is compared with the pre-processed paving operation image I pre (x, y) are concatenated in the channel dimension to obtain the structural feature map F(x, y).

[0017] Preferably, in S3, the local information entropy of the paving operation image is constructed, which can effectively quantify the complexity and information content of the local structure of the paving operation image, accurately identify the texture features of the waterproof board, and have strong robustness to illumination changes and noise interference; a directional entropy gradient field is constructed based on the local information entropy, which can comprehensively cover the structural change features of the paving operation image at all angles, and accurately capture the characteristic response of the laying direction of the waterproof board; a structure manifestation function is constructed based on the disturbance response and the entropy gradient to obtain a structure feature map, and combined with the Gaussian probability distribution disturbance, the sensitivity of feature extraction is enhanced, and through the fusion of the disturbance response and the entropy gradient, the structural features of the waterproof board can be effectively highlighted, and the multi-channel feature fusion improves the integrity of the feature expression.

[0018] Preferably, in S4, the path recognition and path point generation operations are performed on the structural feature map, and the specific method is: using the Otsu algorithm to perform threshold segmentation on the normalized structural response map M′(x, y) to obtain a binary image, using the algorithm to perform skeleton extraction operation on the binary image to obtain a skeleton map, performing pixel traversal on the skeleton map, uniformly extracting path points on the skeleton line according to a preset sampling interval d to form a path point set, performing boundary legitimacy detection on each path point, judging whether the path point is located in the binary structure area, and outputting a path point sequence P that meets the boundary legitimacy, P={P1, P2, ..., P k}, k is the number of output path points.

[0019] Preferably, in S4, the path recognition and path point generation operations are performed on the structural feature map, and adaptive threshold segmentation is achieved through the Otsu algorithm. The optimal segmentation threshold can be automatically determined according to the grayscale distribution of the paving operation image, ensuring the accurate separation of the waterproof board area and the background. The skeleton extraction technology can accurately obtain the central axis characteristics of the waterproof board paving area, while maintaining the integrity of the topological structure and achieving the simplest expression of the path, providing an ideal basic path for robot navigation. In the path point generation link, uniform sampling is performed according to the preset spacing, which not only ensures the reasonable distribution of the path points, but also significantly reduces the amount of data. The validity of each path point is strictly verified through the boundary legitimacy detection mechanism to ensure that the generated navigation path is completely within the operable area.

[0020] Preferably, in S5, a sliding window mean filter is used to perform noise reduction processing on the path point sequence, and the specific method is: The path point sequence is subjected to noise reduction processing by using a sliding window mean filter, the size of the sliding window is set to 3, and each path point P in the path point sequence P is filtered. i , take the previous point P i-1 、Current point P i And the last point P i+1 Averaging is performed to obtain the filtered path point sequence P′.

[0021] Preferably, in S5, the path point sequence is smoothly fitted by B-spline curve fitting to obtain the robot motion path, and the specific method is: A smooth path is generated by B-spline curve fitting. According to the distribution density of the path sequence P′, a portion of points are evenly sampled as control points to obtain the control point sequence ControlPoint. A third-order B-spline fitting is used to construct a B-spline curve based on the control point sequence to obtain the smooth path points. The specific calculation process is as follows: Where, P smooth(t) is the smooth path point generated when the parameter is t, t is the parameter variable on the path curve, n is the total number of control points minus 1, N j，3 (t) is the cubic B-spline basis function corresponding to the j-th control point; As t changes uniformly within the path parameter range, the above formula outputs a series of continuous smooth path points, and the optimized path sequence P is obtained. smooth .

[0022] Preferably, in S5, the path point sequence is subjected to noise reduction processing by sliding window mean filtering, which effectively filters out high-frequency noise and abnormal fluctuations in the path point sequence, while completely retaining the overall direction characteristics of the path. The path point sequence is smoothly fitted by B-spline curve fitting to obtain the robot motion path. The path generated by this method has second-order continuity, which can ensure that the robot moves smoothly without jitter.

[0023] Compared with the prior art, the present invention has the following technical effects: The present invention uses an industrial camera carried by a waterproof board laying robot to collect images of the waterproof board laying area in the tunnel, and performs perspective transformation in combination with the internal and external parameters of the camera, which can effectively eliminate the distortion of the acquired image and obtain the laying operation image under the orthographic perspective; the laying operation image is preprocessed by Gaussian blur, contrast enhancement and brightness normalization, thereby improving the quality of the laying operation image; after the preprocessing of the laying operation image is completed, a directional entropy gradient field is constructed based on local information entropy, and the disturbance response is integrated to generate a structure manifestation function, and the structural feature map of the laying operation image is extracted; the path point sequence for the navigation control of the laying robot is obtained by image skeleton extraction and the path point equidistant sampling method, which can effectively ensure the integrity of the path, and the sliding window mean filter is used to perform denoising on the path point sequence, which can effectively eliminate the noise of the path point sequence, and the path sequence is smoothly fitted by B-spline curve fitting to obtain the robot motion path. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a step diagram of the path recognition method of the waterproof board laying robot provided by the present invention.

[0025] Figure 2 This is a flowchart of the paving operation image preprocessing provided by the present invention. DETAILED DESCRIPTION

[0026] The present invention uses an industrial camera carried by a waterproof board laying robot to collect images of the waterproof board laying area in the tunnel, performs perspective transformation based on the internal and external parameters of the camera, and obtains laying operation images under an orthographic perspective; performs preprocessing operations such as Gaussian blur, contrast enhancement, and brightness normalization on the laying operation images; after the preprocessing of the laying operation images is completed, a directional entropy gradient field is constructed based on local information entropy, and a structural visualization function is generated by fusing the disturbance response to extract the structural feature map of the laying operation image; the path point sequence for the navigation control of the laying robot is obtained through image skeleton extraction and the method of equidistant sampling of path points, the path point sequence is denoised by using sliding window mean filtering, and the path sequence is smoothed by B-spline curve fitting to obtain the robot motion path.

[0027] See Figure 1 As shown, a path recognition method for a waterproof board laying robot based on machine vision in an embodiment of the present application.

[0028] S1. Use the industrial camera carried by the waterproof board laying robot to collect images of the waterproof board laying area in the tunnel. Combined with the internal and external parameters of the camera, a perspective transformation is performed to obtain the laying operation image under the orthographic perspective.

[0029] Furthermore, in S1, the industrial camera carried by the waterproof board laying robot is fixedly installed on the top bracket of the robot, and a fixed shooting angle and height are set so that the camera can completely cover the tunnel laying area. During the operation of the robot, the industrial camera collects images in the tunnel at a set frequency. In order to correct the distortion of the image caused by the tilted perspective, the system pre-completes the internal and external parameter calibration of the camera and obtains the camera imaging model parameters. The collected image is then transformed in perspective based on the calibration parameters, and the original image is mapped from the tilted perspective to the orthographic perspective image to obtain the laying operation image under the orthographic perspective.

[0030] S2. Performing Gaussian blur, contrast enhancement, and brightness normalization preprocessing operations on the paving operation image.

[0031] Furthermore, in S2, the paving operation image is preprocessed, such as Figure 2 As shown, the specific method is: Perform Gaussian blur processing on the paving operation image with a Gaussian kernel size of 5×5 to obtain a blurred paving operation image. Calculate the high-frequency residual map between the paving operation image and the blurred paving operation image. Fuse the high-frequency residual map with the paving operation image. During the fusion, the high-frequency residual map accounts for 0.8, and obtain a high-frequency enhanced paving operation image. The high-frequency enhanced paving operation image is divided into multiple sub-regions of size 64×64. The brightness mean and structural gradient variance of each sub-region are calculated respectively. According to the brightness mean of each sub-region, the brightness normalization operation is performed on each sub-region. According to the structural gradient variance of each sub-region, the contrast adjustment operation is performed on each sub-region. The processed pixels of each sub-region are spliced ​​into a complete image to obtain the preprocessed paving operation image.

[0032] S3. Use the image entropy flow offset response method to extract features from the pre-processed paving operation image, including constructing the local information entropy of the paving operation image, constructing a directional entropy gradient field based on the local information entropy, and constructing a structure visualization function based on the disturbance response and the entropy gradient to obtain a structure feature map.

[0033] Furthermore, in S3, the local information entropy of the paving operation image is constructed, and the specific method is as follows: For the pre-processed paving operation image I pre For each pixel position (x, y) in (x, y), in the neighborhood N(x, y) centered at (x, y), the size of the neighborhood is 5×5, and the distribution probability of the grayscale values ​​of all pixels in the neighborhood is calculated to obtain the local grayscale probability distribution P (x，y) (v), the specific calculation process is: Where v is the pixel grayscale value, v∈[0,255], δ(·) is the indicator function, which counts the number of occurrences of each pixel grayscale value; the local information entropy E(x, y) of the pixel position is calculated based on the local grayscale probability distribution. The specific calculation process is: Where ∈ is a very small integer, ∈=1×10 -5 .

[0034] Furthermore, in S3, a directional entropy gradient field is constructed based on the local information entropy. The specific method is as follows: Define the direction set Θ for detecting direction changes, Θ = {0°, 45°, 90°, 135°, 180°, 225°, 270°, 315°}, and for each direction θ, θ∈Θ, define the unit pixel offset (Δx θ , Δy θ ), for each pixel position (x, y), calculate the information entropy difference F of (x, y) in direction θ θ (x, y), the specific calculation process is: F θ (x, y) = E(x, y) - E(x + Δx θ , y+Δy θ ); The maximum value of the information entropy difference in all directions is taken to obtain the directional entropy gradient response intensity S(x, y) of the pixel. The S(x, y) of each pixel in the paving operation image constitutes a directional entropy gradient field.

[0035] Furthermore, in S3, a structure visualization function is constructed based on the disturbance response and entropy gradient to obtain a structure feature map. The specific method is as follows: For each pixel, the local grayscale probability distribution P (x，y) (v) Using Gaussian distribution, construct a normal distribution with the neighborhood grayscale mean as the center to obtain the local Gaussian probability distribution Q (x，y) (v) P (x，y) (v) with Q (x，y) (v) Perform linear mixing to obtain the perturbation probability distribution P′ (x，y) (v), based on P′ (x，y) (v) recalculating the local information entropy of the pixel area to obtain the perturbed local information entropy E′(x, y), and performing a difference calculation between the original local information entropy E(x, y) and the perturbed local information entropy to obtain the perturbation response value ΔE(x, y); The directional entropy gradient response intensity S(x, y) of each pixel position is multiplied by the disturbance response value ΔE(x, y) to obtain the structural response map M(x, y). The maximum and minimum normalization operation is performed on M(x, y) to obtain the normalized structural response map M′(x, y). The normalized structural response map M′(x, y) is compared with the pre-processed paving operation image Ipr e (x, y) are concatenated in the channel dimension to obtain the structural feature map F(x, y).

[0036] S4. Using the image skeleton extraction and path point equidistant sampling method, perform path recognition and path point generation operations on the structural feature graph to obtain a path point sequence for navigation control of the paving robot.

[0037] Further, in S4, the path recognition and path point generation operations are performed on the structural feature map, and the specific method is: the normalized structural response map M′(x, y) is threshold segmented using the Otsu algorithm to obtain a binary image, the binary image is subjected to a skeleton extraction operation using the algorithm to obtain a skeleton map, the skeleton map is pixel-traversed, and the path points on the skeleton line are uniformly extracted according to a preset sampling interval d to form a path point set, and a boundary legitimacy test is performed on each path point to determine whether the path point is located in the binary structure area, and a path point sequence P that satisfies the boundary legitimacy is output, P={P1, P2, ..., P k}, k is the number of output path points.

[0038] S5. Use sliding window mean filtering to perform noise reduction processing on the path point sequence, and perform smooth fitting on the path point sequence through B-spline curve fitting to obtain the robot motion path.

[0039] Furthermore, in S5, a sliding window mean filter is used to perform noise reduction processing on the path point sequence. The specific method is as follows: The path point sequence is subjected to noise reduction processing by using a sliding window mean filter, the size of the sliding window is set to 3, and each path point P in the path point sequence P is filtered. i , take the previous point P i-1 、Current point P i And the last point P i+1 Averaging is performed to obtain the filtered path point sequence P′.

[0040] Furthermore, in S5, the path point sequence is smoothly fitted by B-spline curve fitting to obtain the robot motion path. The specific method is as follows: A smooth path is generated by B-spline curve fitting. According to the distribution density of the path sequence P′, a portion of points are evenly sampled as control points to obtain the control point sequence ControlPoint. A third-order B-spline fitting is used to construct a B-spline curve based on the control point sequence to obtain the smooth path points. The specific calculation process is as follows: Where, P smooth (t) is the smooth path point generated when the parameter is t, t is the parameter variable on the path curve, n is the total number of control points minus 1, N j，3 (t) is the cubic B-spline basis function corresponding to the j-th control point; As t changes uniformly within the path parameter range, the above formula outputs a series of continuous smooth path points, and the optimized path sequence P is obtained. smooth .

[0041] The above are only preferred embodiments of the present invention. It should be pointed out that those skilled in the art can make several modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A method for identifying the path of a waterproof board laying robot based on machine vision, characterized in that: The following steps are involved: S1. Use the industrial camera onboard the waterproofing board laying robot to capture images of the waterproofing board laying area in the tunnel. Combined with the camera's internal and external parameters, perform perspective transformation to obtain an orthographic view of the laying operation. S2. performing Gaussian blur, contrast enhancement, and brightness normalization preprocessing operations on the paving operation image; S3. Extracting features from the pre-processed paving operation image using an image entropy flow offset response method, including constructing local information entropy of the paving operation image, constructing a directional entropy gradient field based on the local information entropy, and constructing a structure visualization function based on the disturbance response and the entropy gradient to obtain a structure feature map; S4. Using a method of image skeleton extraction and equally spaced sampling of path points, perform path recognition and path point generation operations on the structural feature graph to obtain a path point sequence for navigation control of the paving robot; S5. Use sliding window mean filtering to perform noise reduction processing on the path point sequence, and perform smooth fitting on the path point sequence through B-spline curve fitting to obtain the robot motion path.

2. The method for machine vision-based waterproof board laying robot path recognition according to claim 1, characterized in that: In S2, the paving operation image is pre-processed, and the specific method is as follows: Performing Gaussian blur processing on the paving operation image to obtain a blurred paving operation image, calculating a high-frequency residual map between the paving operation image and the blurred paving operation image, and fusing the high-frequency residual map with the paving operation image to obtain a high-frequency enhanced paving operation image; The high-frequency enhanced paving operation image is divided into multiple sub-regions of equal size. The brightness mean and structural gradient variance of each sub-region are calculated respectively. According to the brightness mean of each sub-region, the brightness normalization operation is performed on each sub-region. According to the structural gradient variance of each sub-region, the contrast adjustment operation is performed on each sub-region. The processed pixels of each sub-region are spliced ​​into a complete image to obtain the preprocessed paving operation image.

3. The method for machine vision-based waterproof board laying robot path recognition according to claim 2, characterized in that: In S3, the local information entropy of the paving operation image is constructed, and the specific method is: For the pre-processed paving operation image I pre For each pixel position (x, y) in (x, y), within the neighborhood N(x, y) centered at (x, y), the distribution probability of the grayscale values ​​of all pixels in the neighborhood is counted to obtain the local grayscale probability distribution P (x，y) (v), the specific calculation process is: Where v is the pixel grayscale value, v∈[0,255], δ(·) is the indicator function, which counts the number of occurrences of each pixel grayscale value; the local information entropy E(x, y) of the pixel position is calculated based on the local grayscale probability distribution. The specific calculation process is: Where ∈ is a very small integer.

4. The method for machine vision-based waterproof board laying robot path recognition according to claim 3, characterized in that: In S3, a directional entropy gradient field is constructed based on the local information entropy. The specific method is: Define the direction set Θ for detecting direction changes, Θ = {0°, 45°, 90°, 135°, 180°, 225°, 270°, 315°}, and for each direction θ, θ∈Θ, define the unit pixel offset (Δx θ , Δy θ ), for each pixel position (x, y), calculate the information entropy difference F of (x, y) in direction θ θ (x, y), the specific calculation process is: F θ (x,y)=E(x,y)-E(x+Δx θ ,y+Δy θ ); The maximum value of the information entropy difference in all directions is taken to obtain the directional entropy gradient response intensity S(x, y) of the pixel. The S(x, y) of each pixel in the paving operation image constitutes a directional entropy gradient field.

5. The method for machine vision-based waterproof board laying robot path recognition according to claim 4, characterized in that: In S3, a structure visualization function is constructed based on the disturbance response and entropy gradient to obtain a structure feature map. The specific method is as follows: for each pixel, the local grayscale probability distribution P (x，y) (v) Using Gaussian distribution, construct a normal distribution with the neighborhood grayscale mean as the center to obtain the local Gaussian probability distribution Q (x，y) (v) P (x，y) (v) with Q (x，y) (v) Perform linear mixing to obtain the perturbation probability distribution P′ (x，y) (v), based on P′ (x，y) (v) recalculating the local information entropy of the pixel area to obtain the perturbed local information entropy E′(x, y), and performing a difference calculation between the original local information entropy E(x, y) and the perturbed local information entropy to obtain the perturbation response value ΔE(x, y); The directional entropy gradient response intensity S(x, y) of each pixel position is multiplied by the disturbance response value ΔE(x, y) to obtain the structural response map M(x, y). The maximum and minimum normalization operation is performed on M(x, y) to obtain the normalized structural response map M′(x, y). The normalized structural response map M′(x, y) is compared with the pre-processed paving operation image I pre (x, y) are concatenated in the channel dimension to obtain the structural feature map F(x, y).

6. The method for machine vision-based waterproof board laying robot path recognition according to claim 5, characterized in that: In S4, path identification and path point generation operations are performed on the structural feature graph, and the specific method is as follows: The Otsu algorithm is used to perform threshold segmentation on the normalized structural response map M′(x, y) to obtain a binary image. The algorithm is used to perform skeleton extraction on the binary image to obtain a skeleton map. The skeleton map is traversed pixel by pixel, and path points on the skeleton line are uniformly extracted according to the preset sampling interval d to form a path point set. The boundary legitimacy test is performed on each path point to determine whether the path point is located in the binary structure area. The path point sequence P that meets the boundary legitimacy is output, P = {P1, P2, ..., P k }, k is the number of output path points.

7. The method for machine vision-based waterproof board laying robot path recognition according to claim 6, characterized in that: In S5, a sliding window mean filter is used to perform noise reduction on the path point sequence. The specific method is as follows: The path point sequence is subjected to noise reduction processing by using a sliding window mean filter, the size of the sliding window is set to 3, and each path point P in the path point sequence P is filtered. i , take the previous point P i-1 、Current point P i And the last point P i+1 Averaging is performed to obtain the filtered path point sequence P′.

8. The method for machine vision-based waterproof board laying robot path recognition according to claim 7, characterized in that: In S5, the path point sequence is smoothly fitted by B-spline curve fitting to obtain the robot motion path. The specific method is: A smooth path is generated by B-spline curve fitting. According to the distribution density of the path sequence P′, a portion of points are evenly sampled as control points to obtain the control point sequence ControlPoint. A third-order B-spline fitting is used to construct a B-spline curve based on the control point sequence to obtain the smooth path points. The specific calculation process is as follows: Where, P smooth (t) is the smooth path point generated when the parameter is t, t is the parameter variable on the path curve, n is the total number of control points minus 1, N j，3 (t) is the cubic B-spline basis function corresponding to the j-th control point; As t changes uniformly within the path parameter range, the above formula outputs a series of continuous smooth path points, and the optimized path sequence P is obtained. smooth .

Images