Mining truck edge detection method based on multi-scale edge membership map

Abstract

The invention relates to the technical field of image processing, in particular to a mining truck edge detection method based on a multi-scale edge membership map. The method comprises the following steps: 1, constructing a multi-scale image pyramid: acquiring a grayscale image as input, performing multi-scale pixel sampling, and performing fuzzy processing by using a Gaussian function to construct a Gaussian image pyramid; 2, gradient calculation: respectively calculating the gradient of the image in the x direction and the y direction by using a Scharr operator; 3, the standard deviation of the local gradient difference is calculated, edge pixels and noise points are further effectively distinguished by calculating the standard deviation of the local gradient difference through the statistical characteristics of the local gradient difference, and therefore the edge detection precision is improved; 4, calculating an edge membership degree; 5, fusing a multi-scale edge membership map; 6, adaptive lagging connection: adopting lagging connection to further process the multi-scale edge membership fusion image so as to improve edge continuity and positioning accuracy; according to the invention, the reliability and safety of automatic loading operation are improved.

Description

Technical Field

[0001] The present application relates to the technical field of image processing, and in particular to a mining truck edge detection method based on a multi-scale edge membership graph. Background Art

[0002] The edge detection of the mining truck is a key technology to realize the autonomous loading operation of the electric shovel and an important guarantee for the construction of a smart coal mine. During the autonomous loading operation, if the edge of the mining truck cannot be accurately detected, it is very easy to cause a collision accident between the excavator bucket and the mining truck due to the bucket position being too low, or the impact of the ore falling due to the bucket position being too high, causing damage to the mining truck; during the autonomous transportation process, accurate and automated detection of the edge of the mining truck is a necessary guarantee for the safe automatic driving of the mining truck and the monitoring of the transportation process.

[0003] However, due to the complex environment of the mining operation site, the actual collected mining truck images have many weak edges and noise points. The existing edge detection algorithm is difficult to effectively distinguish weak edges and noise points, resulting in the problem of missing weak edges or too many miscellaneous edges in the mining truck edge detection image, which affects the accuracy of mining truck edge detection and easily causes collision accidents between the electric shovel bucket and the mining truck. At the same time, the existing edge detection algorithm requires too many parameters to be manually set, and the parameter adjustment process is cumbersome, resulting in a low degree of algorithm automation, which is not conducive to actual promotion and application. Summary of the invention

[0004] The present application provides a mining truck edge detection method based on a multi-scale edge membership graph, which is used to accurately and automatically realize mining truck edge detection and can effectively solve the problems in the background technology.

[0005] To achieve the above objectives, the present application provides the following technical solutions: a mining truck edge detection method based on a multi-scale edge membership graph, comprising:

[0006] (1) Construct a multi-scale image pyramid: Take a grayscale image as input and use a Gaussian function to blur it to construct an image pyramid.

[0007] (2) Calculate the gradient: Use the Scharr operator to calculate the gradient of the image in the x-direction and y-direction respectively. By calculating the gradient features, the background pixels can be effectively distinguished from the edge pixels and noise pixels.

[0008] (3) Calculate the standard deviation of local gradient differences: Utilize the statistical characteristics of local gradient differences and calculate the standard deviation of local gradient differences to further effectively distinguish edge pixels from noise points, thereby improving edge detection accuracy.

[0009] (4) Calculation of edge membership: In order to better distinguish weak edges and noise points and improve the accuracy of edge detection, a new edge membership function is constructed by comprehensively utilizing the standard deviation of the gradient and the local gradient difference.

[0010] Gradient mean μ Gl and the standard deviation of the local gradient difference μ Sl The calculations are as follows:

[0011]

[0012] If G is satisfied at the same time l (i,j)≥μ Gl and S l (i,j)≥μ Sl Two conditions, then the marginal membership E l (i,j) is calculated as follows:

[0013] E l (i,j)=G l (i,j)+S l (i,j).

[0014] If only one of the conditions is met, the edge membership E l (i,j) is:

[0015] E l (i,j)=min(G l (i,j),S l (i,j)).

[0016] If both conditions are not met, the edge membership E l (i,j) is:

[0017] E l (i,j)=0.

[0018] After obtaining the edge membership map, the non-maximum suppression method is used to refine the edge.

[0019] (5) Multi-scale edge membership map fusion: Create a multi-scale edge membership fusion map of size . At each scale, if it is a candidate edge pixel, its associated pixel on the fusion map and the l neighboring pixels of the associated pixel perpendicular to the gradient direction are voted. The number of votes for each pixel on the fusion map is counted. If the number of votes meets the following conditions:

[0020]

[0021] Then the pixel D(i,j) is determined as a candidate edge pixel. Where t represents the number of scale levels of the pyramid image.

[0022] (6) Adaptive hysteresis connection: The multi-scale edge membership fusion map is further processed using hysteresis connection to improve edge continuity and positioning accuracy.

[0023] First, create a fused feature map F of size w×h, where the value of pixel F(i,j) is calculated as follows:

[0024] F(i,j)=G 0 (i,j)+S 0 (i,j).

[0025] The high threshold T is calculated using the Otsu method. h as follows:

[0026] T h =k×Otsu(F).

[0027] Among them, Otsu represents the Otsu method, k∈(0,1) is the adjustment factor, and the default value of k is generally 0.8. h The fused feature map F is thresholded and a residual feature map R of size w×h is created, where the value of pixel R(i,j) is calculated as follows:

[0028]

[0029] Then, the low threshold T is calculated based on the remaining feature map R l After the high threshold T h After thresholding, the remaining feature map R has many weak edges. The maximum entropy threshold method is used to calculate the low threshold T l :

[0030] T l =MET(R).

[0031] Wherein, MET stands for Maximum Entropy Thresholding Method.

[0032] Finally, a high threshold T h and low threshold T l The double thresholds as hysteresis connections are used to process the multi-scale edge membership fusion graph D, thereby obtaining the final binary edge graph E. If D(i,j)>T h , then E(i,j) is determined to be an edge pixel, and E(i,j)=255 is set. l , then E(i,j) is determined as a non-edge pixel and E(i,j)=0. l ≤D(i,j)≤T h, it is necessary to further verify whether there is an edge pixel in the local neighborhood of E(i,j). If so, E(i,j) is determined to be an edge pixel, and E(i,j) = 255 is set. If not, E(i,j) is determined to be a non-edge pixel, and E(i,j) = 0 is set.

[0033] Preferably, when constructing a multi-scale image pyramid, first, obtain a grayscale image I 0 As input, its size is s 0 =w×h. Where S 0 Represented as image I 0 The area of ​​image I 0 The width of image I 0 As the scale level increases, the image size is halved. l , its size is: s l =w / 2 l ×h / 2 l ,l∈{0,1,...,t}.

[0034] The Gaussian function is used as the pyramid generation kernel for image I l The constructed pyramid image is represented as: L = {I l |0≤l≤t}.

[0035] Preferably, the Scharr operator convolution template S in the x-direction and the y-direction x and S y The calculations are as follows:

[0036]

[0037] For image I l , the gradient G along the x direction lx and the gradient G along the y direction ly The calculation is as follows:

[0038] G lx =S x *I l .

[0039] G ly =S y *I l .

[0040] Where * represents the convolution operation. l The gradient magnitude G of (i,j) l (i,j) is calculated as follows:

[0041]

[0042] Gradient direction θ l (i,j) is:

[0043] θ l (i,j)=arctan(G ly (i,j) / G lx (i,j)).

[0044] Preferably, on the gradient map, pixel G l (i,j) in the local neighborhood N l Gradient difference within (i,j) The calculation is as follows:

[0045]

[0046] Local neighborhood N l (i,j) is:

[0047] N l (i,j)={G l (i+m,j+n)|-k l ≤m,n≤k l}.

[0048] Pixel G l (i,j) in the local neighborhood N l The mean gradient difference M within (i,j) l (i,j) is calculated as follows:

[0049]

[0050] Pixel G l (i,j) in the local neighborhood N l The standard deviation S of the gradient difference within (i,j) l (i,j) is calculated as follows:

[0051]

[0052] Preferably, the step of thinning the edge using the non-maximum suppression method is:

[0053] a. Determine the gradient direction: First, you need to calculate the gradient strength and direction at each pixel, which can usually be achieved by applying methods such as the Sobel operator and the Canny edge detector.

[0054] b. Traverse each pixel in the image: Traverse all the pixels in the entire image from left to right and from top to bottom. For each pixel E currently under consideration l (i,j), find its two adjacent points Q along the known gradient direction 1 (x1 ,y 1 ) and Q 1 (x 1 ,y 1 ).

[0055] c. Compare and update pixel values: If E l If the gray value of point (i, j) is greater than or equal to the larger of points Q1 and Q2, P is considered to be a local maximum and its original gray value is retained. Otherwise, the gray value of point P is set to 0, indicating that this is not a valid edge position.

[0056] d. Repeat the above process until the entire image is completed: Continue to perform the same operation on the remaining pixels that have not been set to zero until all pixels have been checked.

[0057] Compared with the prior art, the beneficial effects of the present application are as follows: first, the present application constructs a multi-scale image pyramid based on the original mining truck image to provide multi-scale features for edge detection; second, at each scale, a variety of image features such as gradient and standard deviation of local gradient difference are comprehensively utilized to construct a new edge membership map, which effectively improves the ability to distinguish weak edges and noise points and improves the accuracy of mining truck edge detection; then, a multi-scale edge fusion method based on a voting mechanism is proposed to effectively fuse the multi-scale edge membership map and refine the edge through non-maximum suppression; finally, an automated hysteresis connection method is proposed to obtain a continuous binary mining truck edge map; the mining truck edge detection method based on the multi-scale edge membership map proposed in the present application can accurately and automatically realize mining truck edge detection and effectively improve the reliability and safety of autonomous loading operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is the overall flow chart of this method;

[0059] Figure 2 is a flow chart of the adaptive hysteresis connection method;

[0060] Figure 3 FOM experimental results of mining card images. DETAILED DESCRIPTION

[0061] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0062] See also Figure 1-3,This application provides the following technical solutions: a mining truck edge detection method based on a multi-scale edge membership graph;

[0063] (1) Construct a multi-scale image pyramid:

[0064] In order to better utilize the multi-scale characteristics of the image and improve the integrity and anti-noise ability of edge detection, a multi-scale image pyramid is constructed. Input grayscale image I 0 , whose size is s 0 =w×h. As the scale level increases, the image size is halved. l , its size is:

[0065] s l =w / 2 l ×h / 2 l ,l∈{0,1,...,t};

[0066] And use Gaussian function as the pyramid generation kernel for image I l Perform blur processing. The constructed pyramid image is represented as:

[0067] L={I l |0≤l≤t};

[0068] (2) Calculate the gradient:

[0069] The gradient represents the grayscale difference between an image pixel and its neighboring pixels. Usually, the gradient between background pixels is small, while the gradient between edges and noise is large. By calculating the gradient feature, background pixels can be effectively distinguished from edge pixels and noise pixels. The Scharr operator is often used for gradient calculation in practical engineering applications because of its simplicity, practicality and good effect. The Scharr operator convolution template S in the x-direction and y-direction x and S y The calculations are as follows:

[0070]

[0071] For image I l , the gradient G along the x direction lx and the gradient G along the y direction ly The calculation is as follows:

[0072] G lx =S x *I l ;

[0073] G ly =S y *I l ;

[0074] Where * represents the convolution operation. l The gradient magnitude G of (i,j) l (i,j) is calculated as follows:

[0075]

[0076] Gradient direction θ l (i,j) is:

[0077] θ l (i,j)=arctan(G ly (i,j) / G lx (i,j));

[0078] (3) Calculate the standard deviation of the local gradient difference

[0079] Although background pixels can be distinguished from images through gradient calculation, it is difficult to effectively distinguish edge pixels from noise through gradient maps because the gradient values ​​of edge pixels and noise are usually very small. This can easily cause the loss of effective edges or produce too many miscellaneous edges, thereby affecting the recognition accuracy of the edge of the mining card. In order to solve this problem, the statistical characteristics of local gradient differences are used to further effectively distinguish edge pixels from noise by calculating the standard deviation of local gradient differences, thereby improving the edge detection accuracy.

[0080] On the gradient map, pixel G l (i,j) in the local neighborhood N l Gradient difference within (i,j) The calculation is as follows:

[0081]

[0082] Local neighborhood N l (i,j) is:

[0083] N l (i,j)={G l (i+m,j+n)|-k l ≤m,n≤k l};

[0084] Pixel G l (i,j) in the local neighborhood N l The mean gradient difference M within (i,j) l (i,j) is calculated as follows:

[0085]

[0086] Pixel G l (i,j) in the local neighborhood N l The standard deviation S of the gradient difference within (i,j)l (i,j) is calculated as follows:

[0087]

[0088] (4) Calculate edge membership:

[0089] In order to better distinguish weak edges and noise points and improve the accuracy of edge detection, a new edge membership function is constructed by comprehensively utilizing the standard deviation of the gradient and local gradient difference. Gl and the standard deviation of the local gradient difference μ Sl The calculations are as follows:

[0090]

[0091] If G is satisfied at the same time l (i,j)≥μ Gl and Sl(i,j)≥μ Sl Two conditions, then the marginal membership E l (i,j) is calculated as follows:

[0092] E l (i,j)=G l (i,j)+S l (i,j);

[0093] If only one of the conditions is met, the edge membership E l (i,j) is:

[0094] E l (i,j)=min(G l (i,j),S l (i,j));

[0095] If both conditions are not met, the edge membership E l (i,j) is:

[0096] E l (i,j)=0;

[0097] After obtaining the edge membership map, the non-maximum suppression method is used to refine the edge. Non-maximum suppression refers to comparing the candidate edge pixels E in the gradient direction. l The size between (i, j) and its neighboring pixels is used to determine E l Whether (i,j) is a local maximum. If so, keep the original value unchanged; if not, set it to 0.

[0098] (5) Multi-scale edge membership graph fusion:

[0099] Through the voting fusion mechanism, the edge information of multi-scale images is effectively fused, which improves the integrity of edge detection and the robustness of images of different scales and resolutions.

[0100] Create a multi-scale edge membership fusion graph D of size w×h. At each scale l, if E l (i,j) is a candidate edge pixel, and its associated pixel D(i×2 l ,j×2 l ) and the l neighboring pixels of the associated pixel perpendicular to the gradient direction are voted. The votes are counted for each pixel D(i,j) on the fusion image. If the number of votes v meets the following conditions:

[0101]

[0102] Then the pixel D(i,j) is determined as a candidate edge pixel. Where t represents the number of scale levels of the pyramid image.

[0103] (6) Adaptive hysteresis connection:

[0104] In order to improve edge continuity and positioning accuracy, the multi-scale edge membership fusion map D is further processed by hysteresis connection. Hysteresis connection requires setting high and low thresholds first, and considers pixels with intensity values ​​greater than the high threshold as edge pixels, and pixels with intensity values ​​less than the low threshold as non-edge pixels. For pixels with intensity values ​​between the high and low thresholds, observe their local neighborhood pixels. If there are edge pixels in its local neighborhood, the pixel is considered to be an edge pixel; otherwise, the pixel is considered to be a non-edge pixel.

[0105] Aiming at the problem that the traditional hysteresis connection method needs to manually set high and low thresholds and has low adaptability, an adaptive hysteresis connection method is proposed, as shown in the attached manual. Figure 2 As shown. First, create a fusion feature map F of size w×h, where the value of pixel F(i,j) is calculated as follows:

[0106] F(i,j)=G 0 (i,j)+S 0 (i,j);

[0107] The high threshold T is calculated using the Otsu method. h as follows:

[0108] T h = k × Otsu(F);

[0109] Among them, Otsu represents the Otsu method, k∈(0,1) is the adjustment factor, and the default value of k is generally 0.8. hThe fused feature map F is thresholded and a residual feature map R of size w×h is created, where the value of pixel R(i,j) is calculated as follows:

[0110]

[0111] Then, the low threshold T is calculated based on the remaining feature map R l After the high threshold T h After thresholding, the remaining feature map R has many weak edges. Since the Otsu method has poor detection effect on weak edges, it is easy to cause edge loss, while the maximum entropy threshold method uses information entropy to obtain the optimal threshold and has strong weak edge detection ability. Therefore, the maximum entropy threshold method is used to calculate the low threshold T l :

[0112] T l =MET(R);

[0113] Wherein, MET stands for Maximum Entropy Thresholding Method.

[0114] Finally, a high threshold T h and low threshold T l As a double threshold of hysteresis connection, the multi-scale edge membership fusion graph D is processed to obtain the final binary edge graph E. h , then E(i,j) is determined as an edge pixel, and E(i,j)=255 is set; if D(i,j)<T l , then E(i,j) is determined as a non-edge pixel, and E(i,j)=0 is set; if T l ≤D(i,j)≤T h , it is necessary to further verify whether there is an edge pixel in the local neighborhood of E(i,j). If so, E(i,j) is determined to be an edge pixel, and E(i,j)=255 is set; if not, E(i,j) is determined to be a non-edge pixel, and E(i,j)=0 is set.

[0115] Experimental results analysis:

[0116] In order to evaluate the edge detection effect of mining trucks, the classic FOM and F-Measure are used as experimental performance evaluation indicators.

[0117] As a classic edge detection evaluation index, FOM is widely used to evaluate many advanced edge detection operators. It fully considers factors such as the loss of real edge pixels, edge positioning error, and the number of false edge pixels. FOM is calculated as follows:

[0118]

[0119] Among them, N IRepresents the number of edge pixels in the true edge map, N D Indicates the number of edge pixels in the actual detected edge map, d k It represents the positioning error between the kth edge pixel actually detected and its corresponding pixel in the true edge map. r is the adjustment factor, which is usually 1 / 9. The range of FOM is 0 to 1. The larger the value, the better the edge detection effect.

[0120] F-Measure is the harmonic mean of precision and recall, which is also widely used to evaluate many advanced edge detection operators. For each edge pixel in the actual detected edge map, if the pixel matches the edge pixel in the true edge map, it is called a true positive pixel (TruePositivePixel); if it does not match, the pixel is called a false positive pixel (FalsePositivePixel). For each pixel in the true edge map, if the pixel does not match the pixel in the actual detected edge map, it is called a false negative pixel (FalseNegativePixel). The precision P and recall R are calculated as follows:

[0121] P = TP / (TP+FP);

[0122] R = TP / (TP+FN);

[0123] Where TP represents the number of true positive pixels, FP represents the number of false positive pixels, and FN represents the number of false negative pixels. F-Measure is calculated as follows:

[0124] F Measure =2PR / (P+R);

[0125] By simultaneously using FOM and F-Measure as experimental performance evaluation indicators, the detection rate of the real edge of the mining truck image, the suppression rate of the mixed edges and the positioning accuracy of the edge pixels can be comprehensively reflected, thereby realizing an effective evaluation of the mining truck edge detection effect; in the accuracy experiment, four representative edge detection algorithms, namely method 1, method 2, method 3 and method 4, are used as experimental control methods, and FOM and F-Measure are used as evaluation indicators to verify the performance of the multi-scale adaptive mining truck edge detection method based on edge membership graph proposed in the present invention.

[0126] Among them, method 1 is an algorithm proposed by Han et al., which may be an edge detection method based on structure tensor. Structure tensor is a commonly used tool in image processing to estimate the local structure of an image, so that edges can be detected more effectively. This method usually involves calculating the gradient of the image and using this gradient information to construct a structure tensor to extract edges.

[0127] Method 2 is an edge detection algorithm proposed by Rong et al. that may focus on improving the accuracy and robustness of edge detection. This may include using more complex filters or optimization algorithms to reduce the impact of noise and locate edges more accurately. The specific details need to be consulted in the original literature for more detailed information.

[0128] For method 3, use EdgeDrawing

[11] The algorithm is a heuristic edge detection method that aims to achieve real-time and high-quality edge detection. The ED algorithm achieves edge detection through four main steps: first, use Gaussian filtering to suppress noise; then calculate the gradient magnitude and edge direction map; then extract anchor points; and finally connect the anchor points through intelligent routing. This method is not only different from the traditional binary edge map, but also outputs the edge map in vector form as a set of edge segments, each of which consists of a continuous chain of pixels.

[0129] Method 4 is SBED

[12] The algorithm is a selective bilateral edge detection algorithm. It combines the idea of ​​bilateral filtering to retain edge information and suppress noise by considering both spatial proximity and pixel value similarity. The SBED algorithm may introduce more flexibility and selectivity in the edge detection process to adapt to different image features and requirements.

[0130] The mean values ​​of FOM and F-Measure of each edge detection algorithm in the accuracy experiment are shown in Table 1 below. Through analysis, it can be seen that the edge detection effect of the algorithm proposed in the present invention is better than that of other algorithms. For the mean value of FOM, the algorithm proposed in the present invention is 9.7% higher than method 1, 6.9% higher than method 2, 6.5% higher than method 3, and 7.5% higher than method 4. For the mean value of F-Measure, the algorithm proposed in the present invention is 42.7% higher than method 1, 17.7% higher than method 2, 5.8% higher than method 3, and 6.3% higher than method 4.

[0131] Table 1 Mean values ​​of FOM and F-Measure in accuracy experiments (%)

[0132]

[0133] In order to more intuitively compare the performance of the five edge detection algorithms, the image edge detection results of two typical mining trucks are visualized. Figure 3As shown. Both of the selected mining card images have weak edges to a certain extent. It can be seen from the red marked box that compared with other control methods, the method proposed in the present invention has the best weak edge detection ability, the edge is more complete, and it is very close to the edge map in the true value map. However, other edge detection algorithms have relatively weak detection capabilities for weak edges, and there are different degrees of weak edge loss problems, which affects the continuity of the mining card edge. In addition, it can be seen from the figure that the method proposed in the present invention does not introduce unnecessary noise and miscellaneous edges while ensuring the weak edge detection effect, and the edge positioning is accurate. The detected mining card edge effect is very close to the true value map.

[0134] Experiments have shown that the method proposed in the present invention has high edge detection accuracy, can effectively meet the actual needs of mining truck edge detection, and improve the reliability and safety of the electric shovel's autonomous loading operation. This provides an effective technical guarantee for the realization of the automatic loading function of mining trucks.

[0135] Although the embodiments of the present application have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present application, and that the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A mining truck edge detection method based on a multi-scale edge membership graph, characterized in that: The following steps are involved: (1) Constructing a multi-scale image pyramid: Taking a grayscale image as input, performing multi-scale pixel sampling and using a Gaussian function for blurring to construct a Gaussian image pyramid; (2) Calculate the gradient: The Scharr operator is used to calculate the gradient of the image in the x direction and the y direction respectively; by calculating the gradient features, the background pixels can be effectively distinguished from the edge pixels and the noise pixels; (3) Calculate the standard deviation of the local gradient difference: By using the statistical characteristics of local gradient differences and calculating the standard deviation of local gradient differences, edge pixels and noise points can be further effectively distinguished, thereby improving the accuracy of edge detection. (4) Calculation of edge membership: In order to better distinguish weak edges and noise points and improve the accuracy of edge detection, a new edge membership function is constructed by comprehensively utilizing the standard deviation of the gradient and the local gradient difference. Gradient mean μ Gl and the standard deviation of the local gradient difference μ Sl The calculations are as follows: If G is satisfied at the same time l (i,j)≥μ Gl and S l (i,j)≥μ Sl Two conditions, then the edge membership E l (i,j) is calculated as follows: E l (i,j)=G l (i,j)+S l (i,j); If only one of the conditions is met, the edge membership E l (i,j) is: E l (i,j)=min(G l (i,j),S l (i,j)); If both conditions are not met, the edge membership E l (i,j) is: HAVE BEEN l (i,j)=0; After obtaining the edge membership map, the non-maximum suppression method is used to refine the edge; (5) Multi-scale edge membership map fusion: Create a multi-scale edge membership fusion map of size ; at each scale, if it is a candidate edge pixel, its associated pixel on the fusion map and the l neighboring pixels of the associated pixel perpendicular to the gradient direction are voted; the number of votes for each pixel on the fusion map is counted; if the number of votes meets the following conditions: Then the pixel D(i,j) is determined to be a candidate edge pixel; where t represents the number of scale levels of the pyramid image. (6) Adaptive hysteresis connection: The multi-scale edge membership fusion map is further processed using hysteresis connection to improve edge continuity and positioning accuracy; First, create a fused feature map F of size w×h, where the value of pixel F(i,j) is calculated as follows: F(i,j)=G0(i,j)+S0(i,j); The high threshold T is calculated using the Otsu method. h as follows: T h =k×Otsu(F); Among them, Otsu represents the Otsu method, k∈(0,1) is the adjustment factor, and the default value of k is generally 0.8; use a high threshold T h The fused feature map F is thresholded and a residual feature map R of size w×h is created, where the value of pixel R(i,j) is calculated as follows: Then, the low threshold T is calculated based on the remaining feature map R l ; After the high threshold T h After thresholding, the remaining feature map R has many weak edges. The maximum entropy threshold method is used to calculate the low threshold T l : T l =MET(R); Among them, MET stands for Maximum Entropy Threshold Method; Finally, a high threshold T h and low threshold T l As a double threshold of hysteresis connection, the multi-scale edge membership fusion graph D is processed to obtain the final binary edge graph E; if D(i,j)>T h , then E(i,j) is determined as an edge pixel, and E(i,j)=255 is set; if D(i,j)<T l , then E(i,j) is determined as a non-edge pixel, and E(i,j)=0 is set; if T l ≤D(i,j)≤T h , it is necessary to further verify whether there is an edge pixel in the local neighborhood of E(i,j); if so, E(i,j) is determined to be an edge pixel, and E(i,j)=255 is set; if not, E(i,j) is determined to be a non-edge pixel, and E(i,j)=0 is set.

2. The method for detecting an edge of a mining truck based on a multi-scale edge membership graph according to claim 1, characterized in that: When constructing a multi-scale image pyramid, first, a grayscale image I0 is obtained as input, and its size is s0 = w × h; where S0 represents the area of ​​image I0, w is the width of image I0, and h is the height of image I0. As the scale level increases layer by layer, the image size is halved layer by layer; for an image I0 with a size level of l, l , its size is: s l =w / 2 l ×h / 2 l ,l∈{0,1,...,t}; The Gaussian function is used as the pyramid generation kernel for image I l Fuzzy processing is performed; the constructed pyramid image is expressed as: L = {I l |0≤l≤t}.

3. The method for detecting an edge of a mining truck based on a multi-scale edge membership graph according to claim 1, characterized in that: Scharr operator convolution template S in the x-direction and y-direction x and S y The calculations are as follows: For image I l , the gradient G along the x direction lx and the gradient G along the y direction ly The calculation is as follows: G lx =S x *I l ; G ly =S y *I l ; Where * represents convolution operation; pixel I l The gradient magnitude G of (i,j) l (i,j) is calculated as follows: Gradient direction θ l (i,j) is: θ l (i,j)=arctan(G ly (i,j) / G lx (i,j))。 4. The method for detecting an edge of a mining truck based on a multi-scale edge membership graph according to claim 1, characterized in that: On the gradient map, pixel G l (i,j) in the local neighborhood N l The gradient difference d within (i,j) l ij (m,n) is calculated as follows: Local neighborhood N l (i,j) is: N l (i,j)={G l (i+m,j+n)|-k l ≤m,n≤k l }; Pixel G l (i,j) in the local neighborhood N l The mean gradient difference M within (i,j) l (i,j) is calculated as follows: Pixel G l (i,j) in the local neighborhood N l The standard deviation S of the gradient difference within (i,j) l (i,j) is calculated as follows:

5. The method for detecting an edge of a mining truck based on a multi-scale edge membership graph according to claim 1, characterized in that: The steps of using non-maximum suppression method to refine the edge are: a. Determine the gradient direction: First, you need to calculate the gradient strength and direction at each pixel, which can usually be achieved by applying methods such as the Sobel operator and the Canny edge detector; b. Traverse each pixel in the image: Traverse all the pixels in the entire image from left to right and from top to bottom. For each pixel E currently under consideration l (i,j), find its two adjacent points Q1(x1,y1) and Q1(x1,y1) along the known gradient direction; c. Compare and update pixel values: If E l If the gray value of point (i, j) is greater than or equal to the larger of points Q1 and Q2, P is considered to be a local maximum and its original gray value is retained unchanged; otherwise, the gray value of point P is set to 0, indicating that this is not a valid edge position; d. Repeat the above process until the entire image is completed: Continue to perform the same operation on the remaining pixels that have not been set to zero until all pixels have been checked.

Images