A Research On Lightweight Safety Helmet Detection Algorithm Based On Deep Learning

At present,the requirements of fine management and control of domestic industrial enterprises are increasing day by day,and the on-site safety management is the core content of fine management,in which wearing safety helmet is one of the necessary measures to prevent the body and head injury caused by object strike and fall.In order to realize the management of on-site operators wearing safety helmets,most enterprises have installed video monitoring system,and a few large enterprises have started to use intelligent video monitoring system for safety management.It is a very effective management method to manage the wearing of safety helmet by video monitoring.However,there are still three main problems in the detection of helmet wearing in the existing intelligent video monitoring system.One is that the recognition accuracy of helmet is not high enough,the other is that the recognition accuracy of helmet is low,and the third is that the detection and recognition is easy to be interfered by the environment.Therefore,according to the characteristics of video monitoring in the industrial enterprise operation site,with the idea of combining ease of use and accuracy,through the improvement of YOLO v3 method to complete the safety helmet wearing detection work,to detect the wearing status of the operators’ safety helmet.The detection speed of large convolutional neural network system will be very slow on CPU only computers.The paper proposes an optimization algorithm for helmet wear detection based on YOLO v3.The convolution layer(called Light＿YOLO＿v3))in the basic network darknet-53 of Yolo V3 is replaced by Mobile Net,the depth can be separated and convoluted,which can balance the number of parameters,operation efficiency and recognition accuracy,and can make the safety helmet wear detection of operators reach a faster speed.After training and testing,the mAP value of Light＿YOLO＿V3 does not reach the ideal accuracy.We propose three steps to improve Light＿YOLO＿V3: firstly,aiming at the problem that the safety helmet as a small target in the image can not be recognized due to the low utilization rate of shallow features,a new feature fusion strategy is proposed by using the method of feature fusion of image pyramid in feature extraction to make it meet the requirements of safety helmet detection;secondly,according to the rule of dimension clustering,the safety helmet as a detection target is optimized Third,through further analysis,in order to achieve the universal use of the algorithm,combined with the actual monitoring environment,a multi-scale training method is proposed.Because there is no open data set for helmet detection,this paper builds a set of "Fac Image" data set by collecting factory video monitoring image and network image,which is used for training and testing of neural network.According to the experimental results,the improved light proposed in this Light＿YOLO＿V3 algorithm can keep high detection accuracy while dealing with the task of wearing safety helmet of operators,and meet the requirements of real-time and high-speed detection.