Helmet Wearing Detection Based On Improved YOLO Algorithm

According to statistics,most of the safety accidents in the domestic infrastructure industry are caused by operators not wearing safety helmets.As a protective tool to protect people,safety helmets can protect people’s life safety in dangerous environments.However,due to poor supervision and low safety awareness of personnel,the behavior of not wearing safety helmets occurs from time to time.Based on the deep learning algorithm,this paper designs a scheme of intelligent detection of wearing helmets,which solves the defect of manual detection of wearing helmets.Firstly,aiming at the problem that YOLOV3(You Only Look Once)algorithm is easy to miss detection on targets,this paper proposes G-YOLOV3 algorithm:improve the feature extraction layer of YOLOV3;introduces the idea of Generalized Intersection Over Union(GIOU)instead of the original loss function.Then,aiming at the problem that the detection accuracy improvement effect of G-YOLOV3 algorithm is not obvious,S-YOLOV3 algorithm is proposed:add SPP(Spatial Pyramid Pooling)structure to the backbone network of YOLOV3;The loss function uses Complete Intersection over Union(CIOU).It is analyzed that although S-YOLOV3 algorithm improves the detection accuracy of the model,the overall performance improvement effect is not obvious.In view of this situation,the enhanced version of YOLOV4 algorithm of YOLOV3 algorithm is improved,and T-YOLOV4 algorithm is proposed:CSPDarknet53-Tiny is used to replace the original backbone network;After the output of the backbone network,Ring Fenced Bodies(RFBS)structure is added to increase the receptive field of feature extraction,and the FPN structure is used to fuse the two effective feature layers of the backbone network;K-means clustering algorithm is used to better cover the detection target.Experiments show that the average detection accuracy of YOLOV3 algorithm is 87.68%and the detection speed is 19 frames/s;The average detection accuracy of G-YOLOV3 algorithm is 88.52%,and the detection speed is 17 frames/s;The average detection accuracy of S-YOLOV3 algorithm is 89.74%and the detection speed is 16 frames/s;The average detection accuracy of YOLOV4 algorithm is 92.15%and the detection speed is 13 frames/s;The average detection accuracy of T-YOLOV4 algorithm is 90.36%and the detection speed is 27 frames/s.Through comparative experiments,it can be seen that T-YOLOV4 algorithm meets the requirements of high precision and real-time detection.Using T-YOLOV4 algorithm to detect safety helmets in the construction site is more in line with the actual engineering needs.