Abstract: In order to solve the problem of low detection accuracy of existing helmet detection algorithms for small targets, dense targets and complex environments in construction scenes, an improved target detection algorithm based on YOLOv5, which is recorded as YOLOv5-GBC, is proposed. Firstly, Ghost convolution was used to reconstruct the backbone network, which significantly reduced the complexity of the model. Secondly, BIFPN was used to enhance feature fusion, improved the accuracy of the algorithm for small targets and dense targets. Finally, the Coordinate Attention module was introduced, which can allocate attention resources to key areas, thus reducing background interference in complex environments. In order to verify the feasibility, based on the collected helmet data set, a variety of classic algorithms were selected for comparison, and ablation experiments were carried out to explore the improvement effect of each improved module. The experimental results show that the average precision (IOU=0.5) of the improved algorithm YOLOv5 GBC is improved by 4.9%, reaching 93.6%, and the detection speed reaches 124.3FPS. The model is lighter, the detection capability in dense scenes and small target scenes has been significantly improved ,it meets the requirements of safety helmet detection accuracy and real-time, providing a new method for safety helmet detection in complex construction environments.