CN-121982036-A - Construction site monitoring method and system based on machine vision

CN121982036ACN 121982036 ACN121982036 ACN 121982036ACN-121982036-A

Abstract

The application relates to the field of building safety, in particular to a construction site monitoring method and system based on machine vision. And extracting high-confidence seed points based on the seed point confidence level diagram, combining the seed point confidence level diagram, the local directional diagram and the concrete surface gray level image, judging candidate new growth points of the growth path front points through a multi-criterion growth cost function to complete iterative growth, and processing by a competitive arbitration mechanism based on path backtracking verification when different growth paths meet in the iterative growth process to generate a crack skeleton network. And filtering the crack skeleton network to obtain a purified skeleton, and outputting a concrete surface crack result graph. The application has the effects of effectively distinguishing real cracks from interference textures, reducing false positive rate and improving the accuracy and the robustness of crack skeleton extraction.

Inventors

WANG ZHONG
SUN XIAOXIA
XU YAHUI

Assignees

川纳海仪表(浙江)有限公司

Dates

Publication Date: 20260505
Application Date: 20260409

Claims (8)

1. The construction site monitoring method based on machine vision is characterized by comprising the following steps: preprocessing the obtained concrete surface gray level image of the construction site to generate a seed point confidence coefficient image and a local direction image; Extracting high-confidence seed points based on the seed point confidence level diagram, combining the seed point confidence level diagram, the local directional diagram and the concrete surface gray level image, judging candidate new growth points of the growth path front edge points through a multi-criterion growth cost function to complete iterative growth, and generating a crack skeleton network through competitive arbitration mechanism processing based on path backtracking verification when different growth paths meet in the iterative growth process; filtering the crack skeleton network to obtain a purified skeleton, and outputting a concrete surface crack result graph; The preprocessing method comprises the steps of sending a concrete surface gray level image into a preset multi-scale direction filtering unit for processing, obtaining a modulated multi-scale filtering response through neighborhood contrast modulation operation, inputting the modulated multi-scale filtering response and the concrete surface gray level image into an encoder-decoder network after splicing, generating a fusion weight map through a double-channel attention feature modulation mechanism, carrying out weighted modulation on basic features to obtain fusion features, and sending the fusion features into two parallel output convolution layers to obtain a seed point confidence map and a local directional map.
2. The machine vision based job site monitoring method according to claim 1, wherein the multi-criterion growth cost function obtaining method comprises: Obtaining a local texture dominant direction of a current growth front point based on a local directional diagram, and calculating the deviation degree of the dominant direction and the vector direction of the current growth front point pointing to a candidate new growth point to obtain a direction deviation cost; Based on the gray level image of the concrete surface, obtaining the relative difference degree of the gray level value of the candidate new growth point and the average gray level value of the pixel points grown in the current growth path to obtain the intensity consistency cost; obtaining curvature continuity cost based on the vector direction of the current growth leading edge point pointing to the corresponding precursor point and the change amplitude of the vector direction of the current growth leading edge point pointing to the candidate new growth point, wherein the precursor point is a previous pixel point of the leading edge point on the growth path, the leading edge point is an end pixel point on the grown path, and the candidate new growth point is a pixel point in eight adjacent areas of the leading edge point; Acquiring the seed point confidence coefficient of the candidate new growth point based on the seed point confidence coefficient map to obtain a path seed point confidence coefficient increment; respectively setting weight coefficients for the direction deviation cost, the intensity consistency cost, the curvature continuity cost and the path seed point confidence coefficient increment item, wherein the sum of all the weight coefficients is 1; And multiplying the direction deviation cost, the intensity consistency cost and the curvature continuity cost by corresponding weight coefficients respectively, and summing the multiplied by the path seed point confidence coefficient increment item and the corresponding weight coefficients to obtain the multi-criterion growth cost function.
3. The machine vision based job site monitoring method as set forth in claim 1, wherein the filtering fracture skeleton network comprises: Calculating the second path average curvature change rate and the width change coefficient of each framework branch in the fracture framework network, taking the sum of the global average curvature change rate and the global standard deviation of the second path average curvature change rate as an adaptive threshold value of the second path average curvature change rate, taking the sum of the global average curvature change rate and the global standard deviation of the width change coefficient as an adaptive threshold value of the width change coefficient, and judging that the branch is a pseudo fracture branch with abnormal morphology when the second path average curvature change rate of one framework branch exceeds the corresponding adaptive threshold value and the width change coefficient of the branch simultaneously exceeds the corresponding adaptive threshold value, and filtering the pseudo fracture branch from the fracture framework network.
4. The machine vision based job site monitoring method as set forth in claim 2, wherein the competitive arbitration mechanism for path backtracking verification includes: when candidate new growth points of the two growth paths are the same pixel point, calculating the first path average curvature change rate and the path seed point confidence degree attenuation slope of the two growth paths, multiplying the first path average curvature change rate and the path seed point confidence degree attenuation slope to obtain a judging function value, comparing the size of the judging function values of the two growth paths, and judging that the pixel point belongs to the growth path corresponding to the maximum judging function value.
5. The machine vision based job site monitoring method as set forth in claim 1, wherein the neighborhood contrast modulation comprises: And for each filter response output by the multi-scale direction filter unit, taking a square neighborhood of a preset size of the pixel point as a local window, calculating the ratio of the maximum gradient amplitude to the average gradient amplitude in the local window as a neighborhood contrast sharpening factor, and weighting the filter response through the neighborhood contrast sharpening factor to obtain the modulated multi-scale filter response.
6. The machine vision based job site monitoring method according to claim 1, wherein the extracting high confidence seed points based on the seed point confidence map comprises: calculating the global mean value of the seed point confidence coefficient map and the global standard deviation of the seed point confidence coefficient map, and taking the sum of the global mean value of the seed point confidence coefficient map and the global standard deviation of the seed point confidence coefficient map as a seed point extraction threshold; extracting pixel points with seed point confidence coefficient higher than the extraction threshold value in the seed point confidence coefficient map as candidate seed points; extracting connected domains of candidate seed points, eliminating candidate seed points corresponding to the connected domains with the pixel points less than 3 pixel points, and obtaining the residual candidate seed points as high-confidence seed points.
7. The machine vision-based construction site monitoring method according to claim 1, wherein the preset multi-scale direction filtering unit is composed of a plurality of Gabor filters of different directions and different scales.
8. A machine vision based job site monitoring system comprising a processor and a memory storing computer program instructions that when executed by the processor implement the machine vision based job site monitoring method according to any one of claims 1-7.

Description

Construction site monitoring method and system based on machine vision Technical Field The application relates to the field of building safety, in particular to a construction site monitoring method and system based on machine vision. Background In the machine vision monitoring scene of the concrete structure surface, the core technical pain point is to distinguish the real structural cracks and a large number of similar interference textures of vision, such as template marks, construction joints and the like. In the prior art, the general semantic segmentation model has high false positive rate under the interference, and the root factor is that the extracted features of the general semantic segmentation model have insufficient degree of distinguishing step-type edges from cracks and periodically texture of the interference. In addition, the traditional region growing algorithm only depends on local direction and gray information, is extremely sensitive to direction diagram estimation errors and image noise, is easy to generate problems of growing shake, misleading, incapability of processing branch competition and the like, causes inaccurate generation of a crack skeleton network, and is difficult to meet the requirement of high-precision safety monitoring on a construction site. Therefore, a monitoring method capable of enhancing feature discrimination, having multi-criteria growth constraint and competitive decision mechanism is needed to solve the above-mentioned technical problems. Disclosure of Invention In order to solve the problem of inaccurate generation of the fracture skeleton network, the application provides a construction site monitoring method and system based on machine vision. In a first aspect, the present application provides a machine vision-based construction site monitoring method, which adopts the following technical scheme: The construction site monitoring method based on machine vision comprises the steps of preprocessing an obtained concrete surface gray level image of a construction site to generate a seed point confidence level image and a local direction image, extracting high confidence level seed points based on the seed point confidence level image, combining the seed point confidence level image, the local direction image and the concrete surface gray level image, judging candidate new growth points of a growth path leading edge point through a multi-criterion growth cost function to complete iterative growth, generating a crack skeleton network through competitive decision mechanism processing based on path backtracking verification when different growth paths meet in the iterative growth process, filtering the crack skeleton network to obtain a purified skeleton, outputting a concrete surface crack result image, wherein the preprocessing method comprises the steps of obtaining a modulated multi-scale filter response through sending the concrete surface gray level image into a preset multi-scale direction filter unit, inputting the modulated multi-scale filter response into an encoder-decoder network after splicing the concrete surface gray level image, generating a fusion weight image through a dual-path attention feature characteristic modulation mechanism, weighting and modulating basic features, and sending the fusion weight features into a convolution feature layer to obtain the two-scale feature parallel to the local direction image. The method comprises the steps of obtaining a local texture leading direction of a current growth leading point based on a local directional diagram, calculating the deviation degree of the leading direction and the vector direction of the current growth leading point pointing to a candidate new growth point to obtain a direction deviation cost, obtaining the relative difference degree of a gray value of the candidate new growth point and an average gray value of a growing pixel point of a current growth path to obtain an intensity consistency cost based on a concrete surface gray level image, obtaining a curvature continuity cost based on the change amplitude of the vector direction of the current growth leading point pointing to a corresponding precursor point and the vector direction of the current growth leading point pointing to the candidate new growth point, wherein the precursor point is a leading pixel point of the growth path, the leading point is a tail pixel point of the growing path, the candidate new growth point is a pixel point in eight adjacent domains of the leading point, obtaining a path seed point confidence coefficient increment based on a seed point confidence level diagram, setting weight coefficients for the direction deviation cost, intensity consistency cost, curvature continuity cost and path seed point confidence increment item, respectively, subtracting the weight coefficients from the weight coefficients, and the gain coefficient of the path seed point confidence increment coefficient, respectively, subtra