A Detection Method of Safety Helmet Wearing Based on Centernet

In some construction sites, workers often do not wear helmets and cause safety accidents. In order to prevent the occurrence of safety accidents caused by not wearing safety helmets, we propose a detection method of safety helmet wearing based on CenterNet. Input image into fully convolutional network to obtain a heat map, and the peak of the heat map corresponds to the center of the target. The image features on each peak can predict the width and height of the target frame. The network uses dense supervised learning for training. The inference stage is a single forward propagation network without NMS post-processing. Using video capture from the construction site as a part of the data set. Theoretical analysis and experimental results show that when using detection method of safety helmet wearing based on CenterNet, its recognition accuracy and rate meet the requirements of helmet wearing detection.