Toward Efficient Safety Helmet Detection Based on YoloV5 With Hierarchical Positive Sample Selection and Box Density Filtering
Zhishan Li, Wenqing Xie, Lingzhi Zhang, Shan Lu, Lei Xie, Hongye Su, Weidong Du, Weifeng Hou
Abstract
Object detectors based on convolutional neural networks (CNNs) have been widely deployed in industrial production for safety detection to guarantee the security of workers, and safety helmet detection is one of the most crucial application scenarios. In this article, we first consider the insufficiency of the existing largest open-source helmet detection dataset SHWD and introduce our safety helmet detection dataset which contains various industrial scenarios that are lacking in the former dataset. In addition, we reconsider the sample selection method of the Yolo series and propose a hierarchical positive sample selection (HPSS) mechanism in the training process, which improves the fitting ability of YoloV5. Furthermore, inspired by object detection in continuous frames from videos, we propose a post-processing algorithm based on box density to effectively suppress the appearance of false detection. Under the confidence threshold is 0.1, the combination of the two optimization strategies improves the <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$F1$ </tex-math></inline-formula> -dscore of YoloV5s by 12.47% without increasing any calculation. The dataset will be open-source in the near future.