The UAV target positioning system based on traditional algorithms has low accuracy and is easily interfered by light conditions.Aiming at this problem,a UAV target positioning system based on embedded vision is proposed.The Darknet-19 backbone in the original YOLOv2 is replaced with depthwise separable convolutions,which greatly reduces the model size.The RepVGG block is introduced to extract complex features to improve detection accuracy.The model is deployed to the embedded system and the performance test is carried out.The results show that the model detection accuracy of the improved YOLOv2 algorithm reaches 96.7%,and the detection speed reaches 25 fps,which solves the problem that the traditional algorithm is difficult to deal with illumination changes and has obvious performance improvement.Test on UAV is designed and completed,the test results verified the effectiveness and reliability of the UAV target positioning system.