Aiming at the problems of low detection accuracy and poor effect of safety helmet wearing in heavy industrial sites such as construction industry and construction site, a YOLO ( you only look once ) v7 detection algorithm for safety helmet is proposed. The Convolutional Block Attention Module ( CBAM ) is combined with the neck network of YOLOv7 to reduce the interference of complex background and improve the attention of operators. Then, the detection head of YOLOv7 is replaced by a dynamic detection head DyHead ( Dynamic Head ) to unify multiple attention operations and improve the efficiency of feature fusion, so as to effectively solve the feature fusion problem of small-size target detection. By optimizing the model bounding box regression, the W-IOU ( weighted Intersection over Union ) loss function replaces the original loss function to improve the model training speed and accuracy. The experimental results on the SHWD safety helmet dataset show that the improved algorithm improves the average detection accuracy by 2.2 %, the accuracy rate by 1.1 %, and the recall rate by 1.8 %.The improvement of this paper makes the model more accurate to identify the target, and the detection effect is greatly improved.