Abstract: In order to solve the problems of low safety detection accuracy of existing construction safety detection models in complex environment construction sites and difficult to be deployed in low performance mobile devices, this paper proposes an improved YOLOv8n construction safety detection model CMSFD-YOLO.First, a Channel Prior Convolutional Attention (CPCA) mechanism is introduced into the network backbone to enhance the model's ability to extract construction safety features in complex environments. Convolutional Attention (CPCA) is introduced into the network backbone to enhance the model's ability to extract construction safety features in complex environments. Subsequently, the CMSFD feature pyramid is constructed by combining the Multi Scale sequence Feature fusion (MSF) module and the Triple Feature Encoder (TFE) module, which fully combines the information of multiple types of construction safety targets to optimize the efficiency of multi-scale target feature fusion. Second, the Lightweight Shared Detail-Enhanced Convolutional Detection (LSDECD) detection head is designed to lighten the model and reduce the number of parameters, computation and network model volume, while balancing the detection accuracy. Finally, Focaler-EIoU is introduced as the loss function for bounding box regression to optimize the model convergence speed and regression results. Experiments on the self-built SODA_VOC dataset, and the CMSFD-YOLO model mAP@0.5 it reached 86.2%, are improved by 4.3% compared to YOLOv8n, while the number of parameters, computational volume and model size are only 2266728, 6.5×10⁹ and 4.39 MB, respectively, which are reduced by 24.6% compared to YOLOv8n, 19.7% and 26.2%, respectively, compared to YOLOv8n. Finally, the improved model is deployed to mobile devices for real-world effect testing, which proves that the improved model has good application performance in safety detection tasks in complex construction site environments.