Construction Site Safety Helmet Wearing Detection Method based on Improved YOLOv5

Abstract

Abstract Aiming at real-time monitoring of whether the construction site workers wear helmets correctly according to regulations, we present an improved model which based on the fifth You Only Look Once (YOLOv5) target detection algorithm for helmet wearing detection. Firstly, the clustering algorithm of the network model in the YOLOv5 is optimized to design size of the bounding box. Second, by inducing multi-scale image input to adapt to different image sizes, which enhance the generalization ability of the model. Finally, Complete Intersection over Union (CIoU) instead of original Generalized Intersection over Union (GIoU). CIoU_Loss can entirely consider the distance which are from center point to the length-width ratio between prediction box and ground truth box. The result of final experiment demonstrate that the improved model has a strong identification capability. The detection average precision reach by 92.1%, which can meet can meet the precision demands in actual situations.