Intelligent Site Detection Based on Improved YOLO Algorithm

An improved YOLOv4 algorithm is proposed to improve the detection accuracy of helmets and reflective clothing for safety problems on construction sites that require simultaneous automatic detection of helmets and reflective clothing. Firstly, the clustering algorithm in YOLOv4 network is optimized, and weighted kernel k-means clustering algorithm is used to analyze the dataset, so as to select the anchor box more suitable for small targets and improve the average accuracy and speed of detection; secondly, the Darknet feature network layer inside the YOLO network is optimized, and the feature maps extracted by quadruple downsampling are doubly upsampled, then convolved and fused with the doubled downsampling, and transported to the subsequent network together with quadruple downsampling, eightfold downsampling, and sixteenfold downsampling to achieve the reduction of the probability of missing detection of small targets. Experimental results show that, the average detection accuracy of the improved algorithm is improved by 8.4% when the helmets and reflective clothing are worn at the same time.