A pipeline of stereo vision based visual odometer for unmanned aerial vehicle

Autonomous navigation is of critical importance for unmanned aerial vehicle (UAV). The navigation system of UAV is usually constructed based on Inertial Measurement Unit (IMU) and vision subsystem. The accumulative error in IMU is an inevitable issue. The application of visual information could assist to correct the localization error and make a more robust navigation system. However, the influence from the environment like the illumination variation, lack of textures, occlusion and so on makes it difficult to build a visual odometer. In this study, a pipeline of stereo vision based visual odometer for UAV navigation has been developed. A modified stereo matching method is proposed which used octree to store the map. The edge features in the stereo images have been enhanced which could provide informative clue for stereo matching. The low texture regions which may lead to inaccurate depth measurement have been detected and removed. Constraints like smoothness, consistency and uniqueness are imposed to refine the disparity map. Based on ORB-SLAM2 architecture, a modified stereo visual odometer with quad match is developed. The stereo matching method is used in the stereo VO for correspondence computation and place reidentification. Extensive experiments have verified the efficiency of the stereo VO pipeline.