In safety helmet wear detection, there are problems such as dense targets and occlusion, and existing detection methods perform poorly in terms of accuracy and real-time performance. Existing detection methods often have limitations in terms of both accuracy and real-time performance. To tackle these problems, a lightweight detection model named CA-YOLO is proposed, which is designed to improve the accuracy and real-time performance of detection. Firstly, the backbone network of YOLOv8 is improved using MobileNetv3 network to reduce the number of parameters and computation to improve the detection speed of the network. The DCNv3 module is introduced in the Neck part to improve the extraction efficiency of the model on spatial features. Secondly, a multi-scale feature extraction module and the coordinate attention mechanism module are added to the YOLOv8 network to effectively add global information to enrich feature information to improve the network feature extraction effect. Finally, the CIOU loss is replaced with the Alpha-IOU function, which accelerates the learning process of the target by setting the weight coefficients to further improve the detection accuracy. Experimental results show that compared to the YOLOv8 model and existing classic and novel algorithms, the CA-YOLO model achieves an average detection accuracy of 91.33%, an improvement of 0.54% over the YOLOv8 model. Additionally, The model size and the number of parameters are reduced by 41% and 39%, and the detection speed is increased by 16.9%. Compared to other models, the CA-YOLO model achieves a good balance between accuracy and real-time performance, satisfying the requirements for detecting coal miners" safety helmet wearing.