A Safety-Helmet Detection Algorithm Based on Attention Mechanism

Abstract

Wearing safety helmet is one of the most effective methods to prevent the head injury of construction workers. However, the existing safety helmet detection algorithms based on deep learning mostly have the defects of high false detection rate of similar targets. Therefore, we propose an improved object detection algorithm based on YOLOv3 by integrating attention mechanism, to increase the accuracy of helmet detection. Firstly, due to being combined with the attention mechanism, the ability of expression of the feature graph in the neural network is enhanced, this improves the robustness of the object detection model. Considering the imbalance of samples in existing helmet detection datasets, the loss function was redesigned to ameliorate the imbalance of positive and negative samples, and the accuracy of detection is improved when the targets overlap each other. The experimental results show that our new algorithm improves the mean average precision (mAP) of helmet detection by 6.4% compared with the previous algorithm and has applicability for helmets at different scenes and in different scales.