Research On Mine Personnel Helmet Wearing Detection Method Based On Convolutional Neural Network

In recent years,national policies on intelligent construction of coal mines have been issued one after another,and the coal mining industry is paying more and more attention to the construction of intelligent mine with "intelligence and safety" as the core.Wearing safety helmets is the most basic and important measure to protect personnel safety in coal mine production.With the rapid rise of artificial intelligence,computer vision technology has penetrated into daily production life,and various industries have begun to use target detection technology to meet the corresponding needs.However,underground coal mines have different characteristics from conventional scenes,which leads to the existing helmet detection methods are not ideal in underground coal mine detection,and at the same time,the characteristics of high risk and hardware equipment limitation in underground mines also put forward more stringent requirements on the lightweight and real-time nature of the detection algorithms.Therefore,a coal mine safety helmet detection algorithm with high accuracy,small parameter quantity,and excellent real-time performance is particularly important for the construction of smart mines.In this thesis,two coal mine helmet detection algorithms are proposed for the above-mentioned related problems,respectively,the M-YOLO algorithm to solve the low detection accuracy of existing models and the MSYOLO algorithm to improve the light-weighting of M-YOLO,the specific research content is as follows:(1)Aiming at the problems of existing models that are difficult to detect accurately because of the low brightness and contrast of the coal mine environment and the small size of the target in the image.A reconfigured feature fusion network and loss function for coal mine helmet detection algorithm M-YOLO is proposed.The algorithm improves the recognition accuracy of the network for small targets by using shallow detail information;by adding a spatial attention module to improve the algorithm’s ability to discriminate between target feature information and useless background,allowing the algorithm to focus and enhance useful features,thus improving the algorithm’s ability to focus on safety helmets in complex environments;the loss function is modified so that the training feedback mechanism of the model can more accurately respond to the difference between predictions and reality,thus making it more suitable for training and optimization of safety helmet targets.(2)Aiming at the problems of the existing models are difficult to be installed and used on the edge equipment in coal mines because of their large resource consumption and failing to balance well between the number of model parameters,real-time and detection accuracy.A lightweight detection algorithm MS-YOLO based on improved M-YOLO is proposed,which constructs a new lightweight feature extraction network,and uses the shuffle attention residual bottleneck SCA-Bottleneck in the feature extraction network to achieve the goal of capturing remote dependencies in spatial directions while retaining accurate location information,To some extent,it improves the decrease in detection accuracy caused by replacing lightweight backbone networks;The use of a spatial pyramid pooling-fast structure and a new convolution module in neck network significantly reduces the amount of model parameters and computation.This thesis has built a mine helmet detection dataset and enhanced the dataset to match the coal mine scenario.The algorithm has been tested on the mine helmet detection dataset and the public safety helmet dataset.The experiment of M-YOLO shows that the network performs well on both public and coal mine datasets,providing an improvement basis for MS-YOLO.The MS-YOLO experiment shows that the algorithm model greatly reduces the parameter amount and calculation amount of the model on the premise of ensuring the detection accuracy of the safety helmet,effectively meets the needs of the use of edge computing equipment in the mine,further helping real-time monitoring and rapid linkage of underground safety operations in coal mines.The thesis has 55 figures,18 tables and 88 references.