Lightweight Motorcycle Helmet Detection Based on Improved YOLOv7-Tiny

In response to the current issues of detection models not effectively balancing accuracy and speed, and failing to screen out helmeted pedestrians in traffic environments, we propose a helmet detection model based on an improved YOLOv7-Tiny architecture. Firstly, we introduce the Gather-and-Distribute mechanism in the model's neck to enhance the fusion of shallow features and improve the semantic information of small target helmets. Additionally, by integrating Deformable Convolution v4 (DCNv4) with the CBL module in the backbone network, we enhance the model's ability to extract features from irregular shapes in the feature map, improving its detection capability for occluded and overlapping targets. Furthermore, for small target detection, we incorporate the idea of a shape-IOU loss function into the Normalized Wasserstein Distance, replacing the CIOU loss function in the model. Both CBL-Dcnv and GD modules improve model accuracy without significantly increasing model weight. This ensures a balance between detection accuracy and speed. Finally, we propose a helmet screening strategy to exclude non-riders wearing helmets and accurately detect riders wearing helmets. Experimental results show that the improved model increases the mean Average Precision (mAP) value by eight percentage points while maintaining detection speed within lightweight standards.