The rapid development and application of unmanned aerial vehicles (UAVs) not only bring convenience to the society, but also pose serious threats to public security, personal privacy, and military security. Therefore, rapid and accurate detection of unknown UAV becomes increasingly important. In addition, in UAV detection technology, the method based on machine vision has the advantages of low cost and simple configuration. This paper proposes an optimized YOLOv3 (You Only Look Once version3) based detection and recognition method for low altitude and fast moving UAV. The residual network and multi-scale fusion are used to optimize the network structure of the original YOLO, and the O-YOLOv3 network is proposed. The training and testing are carried out using the real filmed UAV dataset. The experimental results show that the average precision of the optimized method is better than that of the original method, and the detection speed meets the real-time requirement.