Helmet wearing detection algorithm for e-bike riders based on improved YOLOv7-tiny

• Published in: 2023 5th International Conference on Robotics and Computer Vision (ICRCV) pp. 7 - 13
• Main Authors: Tian, Chiheng, Xie, Yu, Xia, Yuansheng, Sheng, Chuangchuang, Chen, Keqiong, Yuan, Hongchun, Hu, Xueyou
• Format: Conference Proceeding
• Language: English
• Published: IEEE 15.09.2023
• Subjects: e-bike; Head; helmet; Mish; Prediction algorithms; Propagation losses; Roads; Robustness; SimAM; SIoU; Training; Transportation; YOLOv7-tiny
• DOI: 10.1109/ICRCV59470.2023.10329130

In order to more accurately and quickly determine whether e-bike riders in the road are wearing helmets or not, a holistic behavioral research method focusing on riding behavior is proposed, and a helmet wearing detection algorithm for e-bike riders is proposed based on the improved YOLOv7-tiny(You Only Look Once version7-tiny). The improved algorithm first uses a smoother Mish activation function to enhance the nonlinear representation in the network. Secondly, a SIoU(SCYLLA-IoU) loss function that takes into account the regression angle of the prediction frames is introduced to solve the phenomenon of the prediction frames wandering around during training. Finally, in conjunction with the parameter-free attention mechanism SimAM(A Simple, Parameter-Free Attention Module), the parameter-free attention feature aggregation module and CBMS module are designed to enhance the network's attention to key feature information and information transfer between deep and shallow networks. The experimental results show that the improved algorithm improves Precision by2.4%, Recall by 1.8%, and mAP (Mean Average Precision) by 1.3% in the homemade e-bike rider dataset compared to the YOLOv7-tiny algorithm. Effectively solves the phenomenon of missed detections and false detections, and improves the detection effect under different traffic conditions.