Abstract: Unmanned aerial vehicle (UAV) mainly relies on monocular camera in target detection and location. The existing technology exists inaccurate target segmentation and low accuracy in location. In order to improve the performance of application, this paper introduces the instance segmentation algorithm SOLO V2, which has excellent performance and eciency, to segment the target accurately. In order to solve the problems of lack of training samples, target rotation and large scale variation, some methods including data enhancement, multi-scale training and network structure optimization are used. By training and testing the actual pictures, the results show that the algorithm can detect the target more accurately, and basically meet the actual needs. Compared with the original SOLO V2, the accuracy and recall rate are both greatly improved. Finally, some pictures are downloaded from the Internet for testing, and many targets can be detected robustly, which further verifies the e ectiveness of the algorithm.