In view of the difficulties faced by the helmet detection model, such as complex background, dense population and overlapping occlusion of targets, a YOLOv8 helmet detection model based on double-layer routing attention and feature fusion is proposed, which integrates the double-layer routing attention mechanism (BSAM) and the bi-directional feature pyramid network (BiFPN). The model integrates channel attention and BiFormer's double-layer routing attention through the BSAM module, and uses coarse-grained areas to filter non critical value pairs to retain the key parts of the routing area, thus enhancing the network's ability to learn the characteristics of the safety helmet; The BiFPN module is introduced into the model. By eliminating the single connected nodes in the network, the connection between the original input nodes and the output nodes in the same layer is increased, and the strategy of transferring deep semantics from top to bottom and merging shallow features from bottom to top is adopted to achieve bidirectional fusion of features and retain the details of the target; The model uses Wasserstein Loss Function (WLoss) to model the boundary box as a two-dimensional Gaussian distribution, calculate the similarity between the two boundary boxes, make the target information focus on the central area of the boundary box, and reduce the sensitivity to small target position deviation. The experimental results show that compared with YOLOv8n baseline model, the improved model achieves significant performance improvement on the SHWD dataset, with recall and accuracy increased by 9.2 and 3.4% respectively to 88.9 and 92.1%. The algorithm performs well in helmet detection in different scenarios and can adapt to complex and changeable environments.