A computing power adaptive belt deviation amount detection method, medium and system
By employing a computationally adaptive dual-branch design in belt misalignment detection, combined with lightweight target detection and a high-precision semantic segmentation model, the problems of low detection accuracy and insufficient robustness in existing technologies are solved, achieving accurate quantification of belt misalignment and reduction of false alarm rate in complex industrial environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENZHEN YUTUO INTELLIGENT CO LTD
- Filing Date
- 2026-07-02
- Publication Date
- 2026-07-31
AI Technical Summary
Existing belt misalignment detection methods suffer from low detection accuracy and insufficient robustness in complex industrial environments. In particular, they are difficult to achieve real-time high-precision detection on low-computing-power equipment and are easily affected by environmental interference.
A method for detecting belt misalignment that is adaptive to computing power is adopted. Through a dual-branch design, the method dynamically switches between different computing power devices. Low computing power devices use a lightweight target detection model to extract edge corner points, while high computing power devices use a semantic segmentation model to extract edge pixel points. By combining bidirectional nearest neighbor pairing and weighted least squares fitting, the accurate fitting of the belt centerline is achieved.
It achieves accurate quantification of belt misalignment in multi-computing scenarios, reduces false alarm rate, improves detection robustness and adaptability, and is suitable for different hardware environments.
Smart Images
Figure CN122492706A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of belt detection technology, specifically to a computing power-adaptive belt misalignment detection method, medium, and system. Background Technology
[0002] Belt conveyor systems are core equipment for material transport in industrial production, and belt misalignment is a common fault in these systems. If not detected and addressed promptly, it can lead to material spillage, belt wear, equipment damage, and even safety accidents. Existing automated belt misalignment detection methods are mainly divided into two categories: 1. Traditional contact-based detection methods: These methods use devices such as belt misalignment switches and roller sensors to detect belt deviation through physical contact. However, this method is susceptible to belt wear and dust accumulation, resulting in low detection accuracy and high mechanical losses, and is gradually being replaced.
[0003] 2. Vision-based non-contact detection methods: One type uses target detection algorithms to identify belt idlers, determining whether belt misalignment has occurred based on the number of idlers within the detection area. This approach suffers from problems such as false alarms due to missed idler detection and inability to quantify the true amount of misalignment, as idlers vary in color and shape, are affected by operating environment factors, and are easily covered by materials. Another type uses threshold segmentation and edge detection operators to extract belt edges, determining belt misalignment based on the positional relationship between the detected edges and the calibration area. However, this approach lacks robustness under common operating conditions such as light interference, dust obstruction, and changes in belt material, easily leading to discontinuous edge extraction and difficulty in accurately locating belt edges, resulting in false alarms.
[0004] Furthermore, existing vision-based non-contact detection methods do not fully consider the varying computing power requirements of equipment in industrial scenarios: one type of solution relies on large-parameter models to pursue edge extraction accuracy, but this requires high hardware computing power and is difficult to implement real-time inference on low-computing-power devices such as edge gateways and embedded terminals on older production lines. Another type of solution adopts lightweight detection models to adapt to low-computing-power devices, sacrificing detection accuracy and robustness, and is prone to edge positioning deviations under complex operating conditions. Therefore, belt conveyor systems urgently need a flexible and robust method for detecting deviations in complex industrial environments. Summary of the Invention
[0005] To address the technical problems existing in the prior art, this invention provides a computing power-adaptive belt misalignment detection method, medium, and system that enables accurate quantification of belt misalignment.
[0006] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows: A computationally adaptive belt misalignment detection method includes the following steps: Acquire a baseline image of the belt during normal operation, calculate the average pixel width of the belt, and obtain the global pixel equivalent; The system's computing power resources are obtained and compared with a preset computing power threshold. If the system's computing power resources are greater than the preset computing power threshold, the left and right edge pixel sets of the belt are extracted from the real-time belt image. Otherwise, the left and right edge corner pixel sets of the belt are extracted from the real-time belt image. Normalize the pixel set of the left and right edges of the belt or the pixel set of the left and right corner edges of the belt into a left pixel set and a right pixel set; Based on the bidirectional nearest neighbor pairing algorithm, edge points in the left point set and the right point set are paired to select valid point pairs that meet the bidirectional matching conditions. By setting point-level and set-level weights, a combined weight is constructed. Weighted least squares fitting is performed on the weighted midpoints of effective point pairs to obtain the real-time centerline of the belt. The distance between the center point of the real-time belt image and the real-time centerline of the belt is then calculated. ; The belt reference centerline is obtained based on the reference image, and the distance between the center point of the reference image and the belt reference centerline is calculated. ; according to , And the actual deviation is obtained from the global pixel equivalent.
[0007] Preferably, the specific process for extracting the pixel set of the left and right edges of the belt from the real-time image of the belt is as follows: A real-time image of the belt is input into a pre-trained belt edge semantic segmentation model, which adopts an encoder-decoder structure. The encoder extracts multi-scale features from the real-time image through multi-layer convolution operations, and the decoder fuses the multi-scale features to output a binary mask matrix of the same size as the real-time image. In the binary mask matrix, the pixel value corresponding to the belt edge region is 1, and the background region is 0. Then, contour detection is performed on the binary mask, and after denoising, the pixel sets of the left and right edges of the belt are extracted.
[0008] Preferably, the specific process for extracting the set of left and right edge corner points of the belt from the real-time image of the belt is as follows: The real-time image of the belt is input into a pre-trained belt edge target detection model. The belt edge target detection model outputs multiple detection boxes. The four corner points of each detection box are taken to construct a Cartesian point set. The Euclidean distance between all point pairs is calculated and sorted. The midpoint of the two edges with the shortest distance is selected, and the gradient and bias pairs are extracted. The pairs are clustered, and the clustering results are mapped back to the original detection boxes to obtain the left and right edge corner point sets of the belt.
[0009] Preferably, the specific process of the bidirectional nearest neighbor pairing algorithm is as follows: For each left edge point in the left point set, find the point with the closest Euclidean distance in the right point set to form a candidate point pair; then for each right edge point in the right point set, find the point with the closest Euclidean distance in the left point set; when a point pair is the nearest neighbor of each other in two matchings, the point pair is identified as a valid point pair.
[0010] Preferably, the point-level weights are set as follows: Continuous edge points: If the distance between adjacent pixels is less than or equal to n1 pixels, it is determined to be a continuous point and the weight is 1.0; Edge interpolation points: adjacent pixel spacing greater than n2 and less than or equal to n3 pixels, with a weight of 0.7; Isolated point: The distance between adjacent pixels is greater than n3 pixels, and the weight is 0.1; Where n1, n2, and n3 are determined according to the image resolution, n1 <n2<n3。
[0011] Preferably, the formula for calculating the combined weights is:
[0012] in, For the first The combined weights of each valid pair of points and The first The left and right weights of each effective point pair It is a regulating factor.
[0013] Preferably, the actual deviation amount The calculation formula is:
[0014] in The local pixel equivalent is calculated using the global pixel equivalent; where A positive value indicates the belt is veering to the right, and a negative value indicates it is veering to the left. The absolute value... This represents the amount of belt misalignment.
[0015] Preferably, local pixel equivalent The calculation formula is:
[0016] in The average pixel width of the belt; Global pixel equivalent; This is a correction item; Intersection The pixel distance between the left and right edge points, where the intersection point The intersection of the line perpendicular to the belt reference centerline and passing through the center point of the reference image with the effective point pair.
[0017] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, the computer program performing the steps of the method described above when run by a processor.
[0018] The present invention further discloses a computing power adaptive belt misalignment detection system, including a memory and a processor connected to each other. The memory stores a computer program, which executes the steps of the method described above when run by the processor.
[0019] Compared with the prior art, the advantages of the present invention are as follows: Multi-computing-power scenario adaptation: It adopts a dual-branch design of object detection model and semantic segmentation model, which can dynamically switch according to the hardware computing power level and the scene: low computing power devices call the lightweight detection model to quickly extract the corner points of the detection box; high computing power devices enable the segmentation model to obtain a high-precision edge point set through pixel-level masking, thereby solving the contradiction between computing power and accuracy in existing methods and realizing full-scene adaptation from low computing power devices to high computing power smart terminals.
[0020] Post-processing logic differentiation: The detection model quickly filters the left and right edge corner point sets by using the Cartesian product of corner points, extraction of the midpoint of the short side, and (gradient, bias) binary clustering; the segmentation model extracts non-zero pixels through a mask, preserving complete edge features of the contour; and then the bidirectional nearest neighbor method is used to achieve accurate matching of the left and right point sets, providing high-quality data support for subsequent fitting and avoiding offset calculation errors caused by edge misalignment.
[0021] Strong anti-interference capability: A dual weighting strategy combining point type weights and overall set weights is proposed. Differentiated weights are assigned to continuous valid points, isolated noise points, and edge interpolation points. At the same time, the set weights are adjusted according to the reliability of the left and right edge point sets. Then, the centerline of the belt is fitted by weighted least squares method, which effectively suppresses the interference caused by noise, occlusion, and edge breakage in industrial scenarios and ensures the stability of the centerline fitting.
[0022] Quantification of true deviation: By introducing a benchmark offset calibration mechanism, the true deviation is directly quantified and output by comparing the position of the fitted centerline with the benchmark centerline in real time, rather than just judging whether there is deviation. This solves the pain points of existing methods that cannot quantify and have a high false alarm rate. It can be directly connected to industrial inspection systems to achieve accurate early warning and correction. Attached Figure Description
[0023] Figure 1 This is a flowchart of the belt misalignment detection method according to an embodiment of the present invention.
[0024] Figure 2This is a schematic diagram of the belt centerline fitting effect in an embodiment of the present invention; where (a11), (a12), and (a13) are the belt centerline fitting effect diagrams of the lightweight detection branch under viewpoint 1, viewpoint 2, and viewpoint 3, respectively; (a21), (a22), and (a23) are the belt centerline fitting effect diagrams of the high-precision segmentation branch under viewpoint 1, viewpoint 2, and viewpoint 3, respectively; and (a31), (a32), and (a33) are schematic diagrams of the calculation of belt pixel deviation under viewpoint 1, viewpoint 2, and viewpoint 3, respectively. Detailed Implementation
[0025] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.
[0026] like Figure 1 As shown, the adaptive belt misalignment detection method provided in this embodiment of the invention includes the following steps: S1. Belt reference image acquisition and pixel global equivalent acquisition As the foundation of the entire detection method, the first step is to install the image acquisition equipment and obtain the baseline parameters. Specifically, a fixed image acquisition camera is installed at an appropriate location on the belt conveyor system. There are no strict restrictions on the camera's installation location, but the shooting range must cover the entire width of the belt and part of the background area to obtain an effective image containing the complete edge of the belt.
[0027] After installing the camera, turn it on to capture a baseline image of the belt in its normal, unbiased state (i.e., without deviation). Here, the image resolution is set to [resolution value missing]. Simultaneously record the actual width of the belt. (Factory-defined parameters).
[0028] On the reference image, corresponding points on the left and right edges of the belt are manually or automatically selected as reference point pairs, and these reference point pairs are processed as follows: First, the reference point pairs are sorted by their horizontal coordinates (if the belt is arranged vertically in the image, the vertical coordinates are used for sorting). Then, distorted areas at the image edges are excluded, and the core area point pairs in the middle 40% are selected. The Euclidean distance of each core point pair is calculated, and the average value is taken to obtain the average pixel width of the belt. Then calculate the global pixel equivalent. as follows:
[0029] This global pixel equivalent A mapping relationship between pixel size and actual physical size was established, providing basic calibration parameters for converting deviation amount to actual physical values in subsequent steps. After completing the above benchmark data acquisition, the system will save these parameters and enter the real-time detection stage.
[0030] S2. Extraction of point sets from the left and right edges of the belt Based on the baseline calibration, the position information of the left and right edges of the belt is extracted from the real-time acquired images. Considering the significant differences in computing power among industrial equipment, this invention designs a computing power-adaptive dual-branch extraction strategy: The system first obtains the current computing power resource status of the hardware platform, and dynamically selects to execute one of the following branches based on the judgment result of "limited" or "sufficient" computing power (the minimum computing power required to run the dual-branch full inference of this invention is taken as the baseline computing power; available computing power ≤ 60% of the baseline computing power is considered limited, otherwise it is considered sufficient). Specifically, branch one is executed when computing power is limited, and branch two is executed when computing power is sufficient: Branch 1. Extraction of left and right edge corners of the object detection model When the system determines that computing power is limited, a lightweight object detection branch is activated. The camera begins capturing real-time images, which are input into a pre-trained belt edge object detection model to obtain object detection results. The belt edge object detection model outputs multiple bounding boxes, each corresponding to a possible belt edge region, represented as follows:
[0031] in: This is the result set output by the belt edge target detection model. This represents the total number of detection boxes within the result set. Represented as the first in the result set There are 4 detection boxes, each consisting of 4 coordinate points. composition.
[0032] For a single detection box Take the Cartesian point set of its four corner points, calculate the Euclidean distance between all pairs of points and sort them, then select the two edges with the shortest distance: ;
[0033] in: These represent the detection boxes. Two different points inside, For detection box Inner Descartes point set, This represents the set sorted by Euclidean distance. This represents the pair of points with the shortest distance. This represents the pair of points with the second shortest distance.
[0034] Take the midpoint of the two sides with the shortest distance. And calculate its corresponding (gradient, bias) pair. ;in: for Midpoint of the point, for Midpoint of the point, Represented as Two points determine the gradient and bias of a line.
[0035] Perform K-Means clustering on the shortest side pairs of all detection boxes:
[0036] in: Represented as the first A collection of categories, This is represented as the cluster center of the current binary tuple; and Corresponding to categories The average gradient and average bias; After clustering, determine if the number of categories c is less than 2 (the left and right edge categories of the belt). If it is less than 2, it indicates that the left and right edges were not completely identified, and the detection fails. If it is greater than 2, the clustered categories c are... Mapping back to the original detection box yields the corner set corresponding to that category. :
[0037] in Corresponding to categories gradient and bias; Through the above processing, even on low-computing-power devices, the corner point set of the left and right edges of the belt can be quickly obtained; Branch 2. Extraction of left and right edge pixels of the object detection model When the system determines that the computing power is sufficient, the high-precision semantic segmentation branch is activated. The camera begins to capture real-time images, which are input into the pre-trained belt edge semantic segmentation model. This model adopts an encoder-decoder structure, where the encoder extracts multi-scale features of the image through multi-layer convolution operations to obtain semantic segmentation results. The decoder fuses the above features and outputs a binary mask matrix of the same size as the input image. In this matrix, the pixel value corresponding to the belt edge region is 1, and the background region is 0. After obtaining the binary mask, contour detection is performed on the binary mask, and after denoising, the pixel sets of the left and right edges of the belt are extracted.
[0038] S3. Valid point-to-point matching Both the corner point set obtained through branch one and the pixel point set obtained through branch two need to be normalized into a unified representation, including the left-side point set. and the right-side point set They are respectively:
[0039]
[0040] in: and These represent the left and right pixels output by the segmentation model, respectively. These represent the total number of pixels on the left and right sides, respectively.
[0041] After obtaining the edge point sets on both sides, the correspondence between the left and right points is established to facilitate subsequent fitting of the belt centerline. This invention employs a bidirectional nearest neighbor pairing algorithm, thereby effectively improving the matching accuracy.
[0042] Specifically, for the left-side point set Each left edge point On the right-hand point set Find the point with the closest Euclidean distance. Form candidate point pairs Then, for the right-hand point set... Each right edge point On the left-hand point set Find the point with the closest Euclidean distance. A match is only valid if a pair of points are each other's nearest neighbors in both matches. The nearest neighbor is ,and The nearest neighbor is also Only then is the point pair considered a valid point pair. .
[0043] This two-way verification mechanism effectively eliminates one-way mismatches caused by noise or edge breaks, ensuring that the point pairs entering the next fitting step have a high degree of spatial correspondence. S4. Weighted Least Squares Median Fitting After obtaining valid point pairs, the core task of this step is to fit the centerline of the belt based on these point pairs. Considering the interference factors such as noise, obstruction, and edge breakage in the industrial environment, this invention innovatively proposes a dual-weighting strategy to improve the robustness and accuracy of the fitting.
[0044] First, define point-level weighting rules, assigning different weights based on the local continuity features of pixels: Continuous edge points: A pixel is considered a continuous point if the distance between it and its neighboring pixels is less than or equal to n1 pixels. It is weighted with 1.0 to ensure that the fitting trend closely matches the real edge. Edge interpolation point: A pixel is located at an edge break, that is, the distance between adjacent pixels is greater than n2 and less than or equal to n3 pixels. The point is used to perform linear interpolation to supplement the broken interval. The weight is 0.7 to balance continuity and accuracy.
[0045] Isolated points: Pixels whose distance from adjacent pixels is greater than n³ pixels are considered isolated noise points and given a weight of 0.1 to reduce the interference of noise on the fitting. Here, n1, n2, and n3 are determined based on the image resolution. Confirmed; n1 <n2<n3。
[0046] Secondly, set set-level weights are defined and adjusted based on the overall reliability of the left and right edge point sets. An adjustment factor is introduced. ， This is used to balance the impact of differences in the number of points on the left and right sides. When , γ ； on the contrary, γ ;in: For the total number of valid point pairs, and The first The left and right weights of a pair of points.
[0047] Finally, the first The combined weight of each valid point pair as follows:
[0048] For the A valid pair of points, its weighted midpoint The calculation method is as follows:
[0049] in: and Represented as point pairs The x and y coordinates, and Represented as point pairs The x and y coordinates, and Represented as point-to-weighted midpoint The x and y coordinates; Based on this, a weighted least squares objective function is constructed, with the optimization objective being:
[0050] in, Represented as the redundancy coefficient of the linear equation, normalized redundancy, taking... ,get:
[0051] By solving this optimization problem, a fitting expression for the belt centerline equation with double weights is obtained:
[0052] in: , , , , , ; S5. Real-time Deviation Calculation This step is divided into two sub-steps: baseline offset calculation and real-time offset calculation. Specifically: Baseline offset calculation: For the baseline image acquired during installation, the effective point set is obtained using steps S2-S4 above. And fit to obtain Equation of the belt reference centerline at time [time] At this point, find the center point of the reference image. and distance for:
[0053] To more accurately reflect the local scale characteristics of the belt's central region, this invention further introduces the concept of local pixel equivalent. A line perpendicular to... And passing through the center point of the reference image straight line Specifically:
[0054] Take a straight line and intersection Calculate the intersection point Pixel distance between left and right edge points Combined with the average pixel width of the belt and global pixel equivalent Calculate local pixel equivalent :
[0055] in This is a minor correction term and is negligible in most cases.
[0056] Real-time offset calculation: For For each real-time image of the belt being captured, repeat steps S2-S4 to calculate the distance from the center point of the real-time belt image to the real-time belt centerline. distance The actual deviation amount The calculation formula is:
[0057] in A positive value indicates the belt is veering to the right, and a negative value indicates it is veering to the left. The absolute value... This represents the amount of belt misalignment.
[0058] like Figure 2 As shown, the geometric relationship between the belt centerline, image center, and orthogonal straight line is illustrated in both the baseline and deviation states, which helps to intuitively understand the calculation principle of deviation. The first two rows of subfigures show the belt centerline fitting effect under different occlusion and breakage scenarios for the lightweight detection branch and the high-precision segmentation branch under different camera mounting angles (angle 1, angle 2, and angle 3); (a11), (a12), and (a13) are the belt centerline fitting effect diagrams for the lightweight detection branch under angle 1, angle 2, and angle 3, respectively; (a21), (a22), and (a23) are the belt centerline fitting effect diagrams for the high-precision segmentation branch under angle 1, angle 2, and angle 3, respectively. Indicates the reference state time of the belt. This represents the average pixel equivalent of the point pairs covered by the gray rectangular region. The belt centerline is represented by the fitted value at the reference state time. It represents the point that passes through the midpoint of the image and is parallel to... Orthogonal lines; take points Pixel distance between left and right edge points Combined with the average pixel width of the belt and global pixel equivalent Calculate local pixel equivalent .
[0059] The last row of sub-images illustrates how the belt pixel misalignment is calculated. Specifically, (a31), (a32), and (a33) are schematic diagrams illustrating the calculation of belt pixel misalignment under viewing angles 1, 2, and 3, respectively. This indicates the moment when the belt is misaligned. This represents the fitted belt centerline at the moment of belt deviation. Represents the intersection of points in the image Orthogonal straight lines. The distance from the center point of the reference image to the reference center line is calculated. And the distance from the center point of the real-time image of the belt to the real-time center line of the belt. The difference between the two yields the pixel deviation, which can then be combined with the local pixel equivalent. The actual deviation can be obtained by conversion.
[0060] The belt misalignment detection method of the present invention first establishes a belt edge detection model and a semantic segmentation model, respectively, and uses the corresponding model to obtain the left and right edges of the belt based on computing power and scene. For the detection model, the Cartesian product of the corner points of the output detection box is calculated, and the midpoint of the short side is selected to obtain a set of (gradient, bias) binary pairs. Then, the set is filtered by mean clustering to obtain the set of left and right corner points of the belt. For the semantic segmentation model, the set of non-zero pixel points of the binary mask of the left and right edges of the belt is extracted. Then, the left and right point sets are matched based on the bidirectional nearest neighbor method, and the weights of continuous points, isolated points, edge interpolation points and the overall weight of the left and right sets are set respectively. The weighted least squares method is used to fit the centerline of the belt. Finally, the real-time offset is compared with the reference offset to obtain the true misalignment. It has the following technical effects: Multi-computing-power scenario adaptation: It adopts a dual-branch design of object detection model and semantic segmentation model, which can dynamically switch according to the hardware computing power level and the scene: low computing power devices call the lightweight detection model to quickly extract the corner points of the detection box; high computing power devices enable the segmentation model to obtain a high-precision edge point set through pixel-level masking, thereby solving the contradiction between computing power and accuracy in existing methods and realizing full-scene adaptation from low computing power devices to high computing power smart terminals.
[0061] Post-processing logic differentiation: The detection model quickly filters the left and right edge corner point sets by using the Cartesian product of corner points, extraction of the midpoint of the short side, and (gradient, bias) binary clustering; the segmentation model extracts non-zero pixels through a mask, preserving complete edge features of the contour; and then the bidirectional nearest neighbor method is used to achieve accurate matching of the left and right point sets, providing high-quality data support for subsequent fitting and avoiding offset calculation errors caused by edge misalignment.
[0062] Strong anti-interference capability: A dual weighting strategy combining point type weights and overall set weights is proposed. Differentiated weights are assigned to continuous valid points, isolated noise points, and edge interpolation points. At the same time, the set weights are adjusted according to the reliability of the left and right edge point sets. Then, the centerline of the belt is fitted by weighted least squares method, which effectively suppresses the interference caused by noise, occlusion, and edge breakage in industrial scenarios and ensures the stability of the centerline fitting.
[0063] Quantification of true deviation: By introducing a benchmark offset calibration mechanism, the true deviation is directly quantified and output by comparing the position of the fitted centerline with the benchmark centerline in real time, rather than just judging whether there is deviation. This solves the pain points of existing methods that cannot quantify and have a high false alarm rate. It can be directly connected to industrial inspection systems to achieve accurate early warning and correction.
[0064] The adaptive computing power belt misalignment detection method provided by this invention can accurately quantify belt misalignment, reduce the false alarm rate of the inspection system, cope with different production and computing power environments, improve equipment adaptability, and has the advantages of flexible computing power adaptation and strong robustness.
[0065] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, the computer program performing the steps of the method described above when run by a processor. The present invention further discloses a computer system including a memory and a processor interconnected thereto, the memory storing a computer program that performs the steps of the method described above when run by a processor. The medium and system of the present invention, corresponding to the methods described above, also possess the advantages described above.
[0066] The present invention can implement all or part of the processes in the methods of the above embodiments, or it can be implemented by hardware related to computer program instructions. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the above method embodiments. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable storage medium includes: any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. The memory is used to store computer programs and / or modules. The processor implements various functions by running or executing the computer programs and / or modules stored in the memory, and by calling data stored in the memory. The memory may include high-speed random access memory, as well as non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital (SD) cards, flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.
[0067] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.
Claims
1. A computationally adaptive belt misalignment detection method, characterized in that, Including the following steps: Acquire a baseline image of the belt during normal operation, calculate the average pixel width of the belt, and obtain the global pixel equivalent; Acquire the system's computing power resources and compare them with a preset computing power threshold; If the system's computing power resources exceed the preset computing power threshold, then extract the left and right edge pixel sets of the belt from the real-time belt image; otherwise, extract the left and right edge corner pixel sets of the belt from the real-time belt image. Normalize the pixel set of the left and right edges of the belt or the pixel set of the left and right corner edges of the belt into a left pixel set and a right pixel set; Based on the bidirectional nearest neighbor pairing algorithm, edge points in the left point set and the right point set are paired to select valid point pairs that meet the bidirectional matching conditions. By setting point-level and set-level weights, a combined weight is constructed. Weighted least squares fitting is performed on the weighted midpoints of effective point pairs to obtain the real-time centerline of the belt. The distance between the center point of the real-time belt image and the real-time centerline of the belt is then calculated. ; The belt reference centerline is obtained based on the reference image, and the distance between the center point of the reference image and the belt reference centerline is calculated. ; according to , And the actual deviation is obtained from the global pixel equivalent.
2. The adaptive belt misalignment detection method according to claim 1, characterized in that, The specific process of extracting the pixel set of the left and right edges of the belt from the real-time image of the belt is as follows: A real-time image of the belt is input into a pre-trained belt edge semantic segmentation model, which adopts an encoder-decoder structure. The encoder extracts multi-scale features from the real-time image through multi-layer convolution operations, and the decoder fuses the multi-scale features to output a binary mask matrix of the same size as the real-time image. In the binary mask matrix, the pixel value corresponding to the belt edge region is 1, and the background region is 0. Then, contour detection is performed on the binary mask, and after denoising, the pixel sets of the left and right edges of the belt are extracted.
3. The adaptive belt misalignment detection method according to claim 1, characterized in that, The specific process of extracting the left and right edge corner point sets of the belt from the real-time image of the belt is as follows: The real-time image of the belt is input into a pre-trained belt edge target detection model. The belt edge target detection model outputs multiple detection boxes. The four corner points of each detection box are taken to construct a Cartesian point set. The Euclidean distance between all point pairs is calculated and sorted. The midpoint of the two edges with the shortest distance is selected, and the gradient and bias pairs are extracted. The pairs are clustered, and the clustering results are mapped back to the original detection boxes to obtain the left and right edge corner point sets of the belt.
4. The adaptive belt misalignment detection method according to claim 1, 2, or 3, characterized in that, The specific process of the bidirectional nearest neighbor pairing algorithm is as follows: For each left edge point in the left point set, find the point with the closest Euclidean distance in the right point set to form a candidate point pair; then for each right edge point in the right point set, find the point with the closest Euclidean distance in the left point set. A pair of points is considered a valid pair if it is the nearest neighbor of the other in two matching processes.
5. The adaptive belt misalignment detection method according to claim 1, 2, or 3, characterized in that, The specific point-level weights are as follows: Continuous edge points: If the distance between adjacent pixels is less than or equal to n1 pixels, it is determined to be a continuous point and the weight is 1.0; Edge interpolation points: adjacent pixel spacing greater than n2 and less than or equal to n3 pixels, with a weight of 0.7; Isolated point: The distance between adjacent pixels is greater than n3 pixels, and the weight is 0.1; Where n1, n2, and n3 are determined according to the image resolution, n1 <n2<n3。 6. The adaptive belt misalignment detection method according to claim 1, 2, or 3, characterized in that, The formula for calculating the portfolio weights is: in, For the first The combined weights of each valid pair of points and The first The left and right weights of each effective point pair It is a regulating factor.
7. The adaptive belt misalignment detection method according to claim 1, 2, or 3, characterized in that, Actual deviation The calculation formula is: in The local pixel equivalent is calculated using the global pixel equivalent; where A positive value indicates the belt is veering to the right, and a negative value indicates it is veering to the left. The absolute value... This represents the amount of belt misalignment.
8. The adaptive belt misalignment detection method according to claim 7, characterized in that, Local pixel equivalent The calculation formula is: in The average pixel width of the belt; Global pixel equivalent; This is a correction item; Intersection The pixel distance between the left and right edge points, where the intersection point The intersection of the line perpendicular to the belt reference centerline and passing through the center point of the reference image with the effective point pair.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program, when run by a processor, performs the steps of the method as described in any one of claims 1-8.
10. A computationally adaptive belt misalignment detection system, comprising a memory and a processor interconnected, wherein the memory stores a computer program, characterized in that, The computer program, when run by a processor, performs the steps of the method as described in any one of claims 1-8.