Helmet Wearing Detection Using Improved Single Shot Multibox Detector

Abstract

In construction sites, safety accidents often occur because construction workers do not wear safety helmets. In order to detect whether construction workers wear safety helmets more timely and efficiently and reduce the incidence of safety accidents, we propose a safety helmet detection method based on a improved SSD algorithm. Our improved SSD algorithm uses four Feature Fusion Modules to fuse high-level features and low-level features to enhance the semantic information of low-level features and improve the ability of the algorithm to detect small-and medium-scale targets. The experimental results show that when the input size is 300, the mAP of our improved algorithm is 2.2 higher than that of the original SSD algorithm on the PASCAL VOC2007 dataset, the mAP is 1.8 higher on our own helmet dataset, and the detection speed is 51 frames per second. Our algorithm meets the real-time requirements of the helmet detection task, and has better detection performance.