Safety helmet detection algorithm in complex scenarios based on YOLOX
Dongsheng Xiang, Baisong Zhu, Xu Dong Wu, Yangfei Ou · 2022
Aiming at the problems that the existing helmet wearing detection algorithms have many application scenarios and rich detection targets. An improved helmet detection method EPP-YOLOX based on YOLOX is proposed. Firstly, the SPP structure of YOLOX is replaced by Enhance Pyramid pooling structure to strengthen shallow feature extraction. Secondly, YOLOX multi-scale fusion structure is added, and the spatial information and semantic information extracted from the attention mechanism image are added to reduce the loss of image details Finally, the experimental results show that the average accuracy of the proposed method is as high as 93.06%, which is 3.28% higher than the original model, while the size of the model parameters is unchanged, and the detection accuracy is also greatly improved. It realizes the high-precision detection of helmet in different scenes such as weak light scene, occlusion scene and multi-target.