Abstract: To address the challenges that traditional detection algorithms face in meeting the requirements of embedded systems, a modified lightweight YOLOv5s detection algorithm is proposed for real-time object detection. The proposed lightweight detection algorithm first replaces the CSP-DarkNet with ShufflenetV2_cssp in the Backbone layer, and combines the DSPPF_CS module to improve frame rate and receptive field. Then, the improved RFBSD structure is introduced in the Neck layer to optimize parameter distribution and enhance object detection capability. The CIoU_SC loss function and scale scaling mechanism are further applied to optimize bounding box regression. Experimental results demonstrate that the number of parameters is reduced by 78.5%, the frame rate is increased by 10 fps, and the accuracy is improved by 1.9%. This showcases the advantages of lightweight design and high performance, providing support for embedded systems.