HRM: An Intelligent Helmet Recognition Model in Complex Scenes

Abstract

This paper presents an intelligent helmet recognition model in complex scenes based on YOLOv5. Firstly, in construction site projects, consider that the photograph which needs to be identified has numerous problems. For example, helmet’s pixels are too tiny to detect, or a large number of workers makes helmets appear densely. A SE-Net channel attention module is added in different parts of the network layer of the model, so that the improved model can pay more attention to the global variables and increase the detection performance of small target information and dense target information. In addition, this paper constructs a helmet data set based on projects and adds training samples of dense targets and long-range small targets. Finally, the modified mosaic data enhancement reduces the influence of redundant background on the model and improves the recognition accuracy of the tiny target. The experimental results show that in the project, the average accuracy of helmet detection reaches 92.82%. Compared with SSD, YOLOv3, and YOLOv5, the average accuracy of this algorithm is improved by 6.89%, 8.28%, and 2.44% and has strong generalization ability in dense scenes and small target scenes, which meets the accuracy requirements of helmet wearing detection in engineering applications.