For the existing helmet wearing detection algorithm for dense targets and small targets having the phenomenon of missed detection, many parameters, large computation, and not suitable for deployment in the embedded device side and other problems, this paper proposes an improved YOLOv5 helmet wearing detection algorithm YOLOv5-Q. Firstly, on the original network 80×80 feature map, the 2 times up-sampling operation is made to form a 160×160 feature map, and the new feature map fuses the three-layer feature information of the original model to form a four-scale detection, which improves the detection accuracy of dense targets and small targets. Secondly, the feature extraction is achieved by replacing the original YOLOv5 backbone network with a lightweight GhostNet, which reduces the parameters of the network and can be ported to the embedded devices for target detection. Finally, the attention mechanism CA is added to boost the weight of important information in the feature map and suppress the weight of non-relevant information, thus improving the accuracy of the model. The experimental results show that the model size of YOLOv5-Q is 26.47 MB, the number of parameters is 12 696 640, and the accuracy is 0.937. Compared with YOLOv5, the YOLOv5-Q algorithm reduces the number of parameters by 39.12%, the model size by 37.2%, but the accuracy by only 1.2%. The YOLOv5-Q algorithm increases the detection accuracy of small targets in dense environments and meets the requirements for deployment on the embedded side.