AC-YOLO: A Safety Helmet Detection based on YOLOX
Yusong Lin, Mengdi Liu, Cong Yang, Shuang Li, Weixing Zhang · 2022
Aiming at the problems of occlusion, low resolution, many small targets and complex background in the existing scene images, which lead to the low accuracy of helmet wearing detection, a new helmet wearing detection algorithm AC-YOLO is proposed. Firstly, in order to solve the problems of limited size, appearance and geometric features of small targets in the existing datasets, a new helmet data set is constructed in this paper. Based on the original YOLOX model, the AC-YOLO model adds the small object detection head, and the original three-scale feature layer is increased to four-scale feature layers. In addition, the AC-YOLO model adds the AC-SR module and the CBAM attention module on the basis of the original YOLOX algorithm, which can extract as many foreground features as possible while fusing multi-scale features. The experimental results show that, compared with the original YOLOX model, the mAP of this algorithm is improved by 1.75%, the precision of wearing helmet is 95.83%, and that of not wearing helmet is 93.45%, which is improved by 2.22% and 1.24% respectively.