Aiming at the navigation problem of unmanned aerial vehicle (UAV) in the process of aerial recovery, a technology of target detection based on deep learning and pose estimation based on binocular vision is proposed. A visual navigation system for aerial recovery is designed, which improves the detection accuracy and speed in the recovery process by improving the original target detection algorithm YOLOv3 framework. The 3D pose of the feature points is calculated by the binocular vision system, and the relative position information of the UAV and the recovery drogue center is returned. The experimental results show that the average accuracy of the improved algorithm is 3.2% higher than that of YOLOv3, and the detection speed is increased to 73 FPS, which shows that the detection speed is significantly improved. The pose calculation accuracy of the binocular vision algorithm is high, and both of them meet the requirements of accuracy and real-time of navigation system.