An Efficient Helmet Wearing Detection Method Based On YOLOv7-Tiny.

In high-risk construction environments, ensuring safety is paramount. Accidents, often resulting in casualties, underscore the importance of wearing helmets correctly. To tackle the challenges posed by small, traditional targets and the complexity of such scenarios, this paper proposes an efficient helmet detection method based on improved YOLOv7-Tiny. The method optimizes the feature extraction layer network for the detection effect of small targets; and improves the feature fusion method of PANet by adding a cross-layer densely connected network and proposing an H-Dense feature fusion network, which enhances the information exchange between multi-scale features while reducing the number of parameters in the network; in addition, a lightweight upsampling operator is introduced CARAFE, which improves the upsampling process and thus improves the quality of feature fusion. Experimental results on the public SHWD dataset reveal significant improvements. Compared to SSD and YOLOv4-Tiny, the improved algorithm achieves a performance increase of 5.6% and 14.1%, respectively. In comparison to the original algorithm, there is a 21.9% reduction in parameters, while the mAP improves by 1.5%. These outcomes affirm that the proposed method effectively fulfills the safety helmet detection requirements in real-world scenarios.