A Study on Lightweight Algorithms for Detecting Helmet-Wearing in Electric Scooters

In urban transportation, electric bicycles serve as a primary mode of non-motorized transport, leading to increasing concerns about traffic safety, especially the helmet-wearing habits of riders. Helmets are crucial for rider safety, effectively reducing injuries in traffic accidents. Existing helmet detection algorithms struggle with small objects, occluded targets, and complex environments. To address these issues, we propose a helmet-wearing detection algorithm based on YOLOv8n. We enhance the model's feature extraction by incorporating the BiFormer attention mechanism before the first detection head, P3, improving the detection of small objects and overall performance. Additionally, in the Neck part of YOLOv8n, we replace traditional convolution modules with the lightweight G-GSConv module for efficient feature aggregation without sacrificing performance. Building on G-GSConv, we introduce the G-Ghost-Bottleneck layer, using a one-shot aggregation method to design the cross-level part network (GSCSP) module, G-GhostCSP-Bottleneck, replacing the original three CSP modules to reduce model parameters. We further integrate the BSAM attention mechanism based on the Transformer architecture to enhance the detection of dense small objects, thereby improving model accuracy. We also refine the loss function using the MDPIOU method to enhance detection flexibility and robustness. Experimental results on the SHWD dataset show that our improved model achieves an average precision of 88.9%, a 3.6% improvement over the original model. The proposed algorithm effectively improves helmet-wearing detection accuracy in complex traffic scenarios, demonstrating its potential for enhancing urban traffic safety management.