Aiming at the leakage and misdetection problems of electric vehicle helmet wearing detection in the presence of occlusion, dense vehicles and the complex scene of multiple people in one vehicle, an improved algorithm applied to electric vehicle helmet wearing detection is proposed on the basis of YOLOv5s. A GBC3 module improved by recursive gating convolution is designed to replace the C3 module in the network backbone and feature fusion layer (Feature Pyramid Networks, FPN), strengthen the spatial information interaction of neighbor features, and improve the feature extraction and feature fusion capabilities of the network. Secondly, Non-Parametric Attention Mechanism (A Simple, Parameter-Free Attention Module for Convolutional Neural Networks, SimAM) is added to the backbone and feature fusion network to adjust the attention weight of different regions in the feature map and pay more attention to important targets. Finally, the WIOU loss function is introduced to optimize the prediction box regression and improve the target localization ability of the model. The experimental results on the self-made electric vehicle helmet dataset show that the mAP of the improved model reaches 97.3% under the premise of only adding fewer parameters, which is 3.2 percentage points higher than that of YOLOv5s, and the detection speed is 87.1FPS, which improves the problem of false detection and missed detection, and still has high real-time performance, which is more suitable for helmet wearing detection of electric vehicle drivers.