In order to solve the problems of low detection accuracy, large environmental interference factors, and difficult deployment in mobile devices with average performance of the existing detection algorithms for hard hats and reflective clothing on small targets and complex weather, an improved detection algorithm for YOLOv8 safety equipment, YOLOV8-DSI, was designed and implemented. Firstly, the DR-SPPF module based on residual idea and parallel cavity convolution is designed to further expand the receptive field without loss of image resolution, and significantly improve the precision of complex weather detection. Secondly, ST-BiFPN is designed in the feature fusion stage to further reduce the number of model parameters and achieve efficient multi-scale feature fusion. Finally, Inner-ShapeIoU loss function is introduced to make bounding box regression more accurate and enhance the detection effect. Compared with the baseline model mAP50 and MAP50:95, the self-built data set increased by 2.1% and 4.7% respectively, while the model parameter number was only 2.4 M and the calculation amount was only 7.3 G, which decreased by 10.9% and 20.0% respectively. Finally, the improved model was deployed to the edge device of Jetson Orin Nano. The actual operation on the development board proved that the improved model of YOLOv8 was effective and applicable in complex scenarios.