Research on an improved yolov5s algorithm for detecting helmets on construction sites
Qi Liu, Fenggang Han · 2023
Aiming at the situation of irregular wearing of safety helmets on construction sites, a work on the detection of the wearing of safety helmets on construction sites is proposed to improve the yolov5s (You only look once) algorithm. To improve the problem that a large number of small targets are difficult to accurately identify due to dense construction site targets, combined with the structured weighted bidirectional pyramidal feature network (BiFPN) to improve the process of feature merging; in the Yolov5s feature extraction network in the introduction of the CBAM attention mechanism of the convolution module to obtain more feature details; using SIoU instead of CIoU as prediction box regression loss makes the network training and inference process faster and more accurate. The experimental findings show that the model with improved YOLOV5S helmet detection has higher detection precision compared to the YOLOV5S model, with MAP increased by 6.6 percentage points, recall increased by 8 percentage points and accuracy by 0.2 percentage points, making the improved algorithm more satisfactory for the helmet-wearing detection challenge.