Adaptive Scene Change Safety Helmet Wearing Detection Algorithm

Wearing a safety helmet is a basic measure to reduce or avoid accident losses. However, factors such as changes in lighting and target occlusion can cause instability in various test result indicators. To address this issue, a helmet-wearing detection algorithm based on adaptive scene changes is proposed. Taking YOLOv8n as the baseline, the SE attention mechanism is embedded in the basic unit C2f to adjust channel weights, capturing more details and enhancing the ability for multi-scale feature extraction and fusion. The BiFormer attention mechanism is embedded at the front end of the detection head, utilizing dual-layer routing attention to improve target feature capture, enhance detection head performance, and strengthen the algorithm's adaptive scene change capability. The baseline algorithm's loss function, CIoU, is replaced with Inner IoU, focusing more on overlapping areas within the bounding box. This improves detection accuracy, accelerates convergence, and enhances generalization ability. Additionally, the Lion optimizer is used to optimize the algorithm, improving gradient descent performance, reducing memory consumption and computational complexity, and increasing detection speed. Ablation experiments and comparative experiments on the self-made Adaptive Scene dataset show that, compared with the baseline algorithm, Precision, Recall, and mAP are improved by 3.1%, 5.3%, and 4.4 %, respectively, demonstrating the superiority of the proposed algorithm.