CN-122024179-A - Site operation safety wearing detection method based on improved YOLOv algorithm

CN122024179ACN 122024179 ACN122024179 ACN 122024179ACN-122024179-A

Abstract

The invention discloses a field operation safe wearing detection method based on an improved YOLOv algorithm, which relates to the technical field of computer vision and deep learning algorithms, and comprises the following steps of embedding a C3k2-R module and a DSConv module into a backbone network based on YOLOv, replacing an original downsampling module with an RFD double-branch downsampling module, optimizing a neck network with an improved PAN-FPN, replacing an original detection head with a dynamic detection head to form an improved YOLOv algorithm, and detecting a remote shielding safe wearing problem by utilizing the improved YOLOv11 algorithm under a high-risk operation scene. According to the invention, through multi-module collaborative optimization, comprehensive breakthrough is realized on the safety wearing detection performance and engineering practicability, and the small target detection precision is remarkably improved.

Inventors

ZHU WENQIANG
HU YIWEI
PI YU
CHI XINYU
MAO GE
LI JIE
HAN YUE

Assignees

武汉纺织大学

Dates

Publication Date: 20260512
Application Date: 20260413

Claims (8)

1. The field operation safety wearing detection method based on the improved YOLOv algorithm is characterized by comprising the following steps of: based on YOLOv, a C3k2-R module and a DSConv module are embedded in a backbone network, an original downsampling module is replaced by an RFD double-branch downsampling module, an optimized neck network is an improved PAN-FPN, an original detection head is replaced by a dynamic detection head, and an improved YOLOv algorithm is formed; under the high-risk operation scene, the remote shielding safety wearing problem is detected by utilizing the improved YOLOv algorithm.
2. The field operation safety wearing detection method based on the improved YOLOv algorithm according to claim 1, wherein the method comprises the following steps: when YOLOv is improved, a 'multi-scale receptive field self-adaption + space attention weighting' double mechanism is constructed by embedding a RFCBAMConv module into a YOLOv original C3k2 module, so that the problem of insufficient extraction of small-size safety wearing features is solved.
3. The field operation safety wearing detection method based on the improved YOLOv algorithm according to claim 2, wherein: When YOLOv is improved, a dynamic snake-shaped convolution module is introduced, and the target outline is fitted by dynamically adjusting the convolution kernel offset, so that the irregular morphological conditions including safe wearing inclination, edge breakage and local shielding are aimed at.
4. A field operation safety wearing detection method based on a modified YOLOv algorithm according to claim 3, wherein: when YOLOv is improved, an RFD double-branch downsampling module is utilized to replace a YOLOv original stride convolution downsampling module, a 'main branch efficient compression and auxiliary branch lossless transfer' double-branch structure is adopted, and small target feature loss is reduced.
5. The field operation safety wearing detection method based on the improved YOLOv algorithm according to claim 4, wherein: when an RFD dual-branch downsampling module is used, the RFD dual-branch downsampling module consists of a main branch, an auxiliary branch, and feature fusion, wherein, Dividing an input feature map into 4 groups according to channels by using main branches as grouping convolution, extracting local features from each group through 3×3 convolution, forming depth separable convolution by 3×3 depth convolution and 1×1 point convolution, further compressing the feature map size by step length=2, adopting 2×2 pooling cores to screen key features if the maximum pooling is achieved, and outputting the result with the size of 1/2 of the original feature map; The auxiliary branches are space-to-depth convolutions, and space dimension information is transferred to channel dimensions; feature fusion is that the double-branch features are spliced in the channel dimension, the channel number is compressed to the original dimension through 1X 1 convolution, and a downsampling feature map retaining key semantic information is output.
6. The field operation safety wearing detection method based on the improved YOLOv algorithm according to claim 5, wherein: When YOLOv is improved, an MSDA module and a CSPF module are introduced on the basis of the original PAN-FPN, so that fusion of shallow position information and deep semantic information is enhanced, and feature confusion in a dense scene of equipment is solved.
7. The field operation safety wearing detection method based on the improved YOLOv algorithm according to claim 6, wherein: When YOLOv is improved, the original decoupling detection head is replaced by DyHead, and the scale perception, the space perception and the channel perception are integrated, so that the safety wearing classification and positioning precision are enhanced.
8. A field operation safety wearing detection system based on a modified YOLOv algorithm for realizing a field operation safety wearing detection method based on a modified YOLOv algorithm as defined in claim 1, comprising: the data acquisition module is used for acquiring high-risk operation scene images; The safety detection module is used for detecting the long-distance shielding safety wearing problem existing in the high-risk operation scene image by utilizing the improved YOLOv algorithm, wherein based on YOLOv, a C3k2-R module and a DSConv module are embedded in a backbone network, an original downsampling module is replaced by an RFD double-branch downsampling module, a neck network is optimized to be an improved PAN-FPN, an original detection head is replaced by a dynamic detection head, and the improved YOLOv algorithm is formed.

Description

Site operation safety wearing detection method based on improved YOLOv algorithm Technical Field The invention relates to the technical field of computer vision and deep learning algorithms, in particular to a field operation safety wearing detection method based on an improved YOLOv algorithm. Background In high-risk operation scenes such as building construction, mining, chemical production and the like, safety wearing detection is a key link for constructing a safety production defense line, guaranteeing life safety of first-line operators, preventing safety accidents such as object striking and high-altitude falling, casualty accidents caused by incorrect wearing are one of main components of accidents in high-risk industries, and therefore the high-efficiency and accurate safety wearing detection technology has important practical significance for reducing the accident rate. Along with the rapid iteration of the computer vision technology and the deep learning algorithm, the automatic safe wearing detection scheme based on the target detection algorithm gradually replaces the traditional low-efficiency mode depending on manual visual inspection, the traditional manual inspection is not only required to put in a large amount of labor cost, but also has the problems of inspection blind areas, subjective judgment errors and the like, and the automatic scheme can realize real-time detection and abnormal early warning through high-definition cameras and edge computing equipment deployed on an operation site. The YOLO series algorithm adopts an end-to-end detection architecture, has high reasoning efficiency and lightweight deployment characteristics, and becomes a mainstream choice in the field of industrial safety detection. When the existing YOLOv-based detection technology is applied to safe wearing detection, a plurality of core problems still face, and the problems are particularly prominent in high-risk operation scenes such as building construction, mining, chemical production and the like, and directly influence detection precision and safety early warning reliability. Aiming at the problem of insufficient extraction of small-target safe wearing features, in an actual operation scene, part of safe wearing presents remarkable small-size features due to long shooting distance and serious shielding, the pixel size of most small-target safe wearing is only 20 multiplied by 20 to 32 multiplied by 32, a C3k2 module in a YOLOv original model adopts a receptive field design with fixed size, the default convolution kernel size and receptive field range of the receptive field design are more adaptive to a middle-size and large-size target, and fine-size features of small-size safe wearing are difficult to accurately capture. On one hand, the fixed receptive field cannot focus on a small target area, the safety wearing features are easy to be confused with background information, and on the other hand, key features such as edge contours, surface textures and the like of small-size safety wearing are easy to be diluted under the fixed receptive field, so that a model cannot effectively distinguish 'small-size safety wearing' from 'background interferents', and finally, the omission ratio of the small-size safety wearing is higher. In the long-distance monitoring scenes such as mine surface mining areas, the omission rate is even higher, and the safety of unrecognized unworn personnel is seriously threatened. In the actual operation process, the safety wearing is in an irregular form usually due to personnel operation habit, operation action or accidental collision, for example, the safety wearing inclination angle reaches more than 30 degrees when a construction worker bends over to operate, the safety wearing edge is mechanically scratched to cause damage when a mine worker shuttles between devices, the safety wearing part is shielded by a tool box when the chemical worker carries tools, a YOLOv original model adopts a rectangular convolution kernel with a fixed shape, a convolution kernel grid cannot fit the outline of the irregular object, for the inclined safety wearing, the fixed convolution kernel easily brings the background area outside the cap peak into a feature extraction range to cause feature pollution, for the edge damage safety wearing, the fixed convolution kernel is difficult to identify the outline of a cap body with incomplete, the safety wearing object is easily misjudged, for the local shielding safety wearing, the fixed convolution kernel cannot avoid the focusing of the shielding area to the non-shielded cap body feature, and finally the feature extraction is insufficient. According to the actual measurement data, the YOLOv original model has higher misjudgment rate on irregular form safety wearing, and seriously influences the accuracy of the detection result. Aiming at the problem of feature attenuation caused by downsampling, YOLOv is to balance calculation efficiency and feature