AMME-YOLOv7: Improved YOLOv7 Based on Attention Mechanism and Multiscale Expansion for Electric Vehicle Driver and Passenger Helmet Wearing Detection
Hongbin Zhang, Aimin Xiong, Liyue Lai, Chuangzhao Chen, Jingfeng Liang · 2023
To address the current problems of low accuracy of small target detection in electric vehicle helmets and imperfect detection schemes for drivers and rear passengers, in this paper, we propose an improved YOLOv7 based on hybrid attention mechanism and multi-scale extension, named AMME-YOLOv7. The model introduces Convolutional Attention Module (CBAM) and Coordinated Attention Module (CA) to improve the accuracy of YOLOv7 by YOLOv7's Backbone and Head to improve accuracy by inserting these two attention modules to better focus on key information. A helmet detection system for electric vehicle drivers was also built by adding multi-scale feature fusion detection and combining it with a dense connectivity network to improve feature extraction. Validation on a self-built EV helmet fitting dataset showed that the AMME-YOLOv7 algorithm improved the average accuracy by 8.9% and the recall by 2.7% over the original YOLOv7 algorithm. The experimental results show that the proposed improved YOLOv7 algorithm better meets the requirements for the detection accuracy of non-motorised vehicles and their helmets in practical situations, and reduces the incidence of motorised vehicle accidents to a certain extent.