关键词: 视觉里程计, ORB, 混合运动估计, 稀疏光束平差法, Kinect

Abstract: Due to the limit of depth information range of RGB-D camera and their vulnerability to noise, in order to improve the accuracy of the RGB-D visual odometry, a method of visual odometry with hybrid motion estimation model based on 3D-3D and 3D-2D is proposed. Firstly, using the 3D-3D model based on RICP algorithm（RANSAC-ICP）, combined with the 3D-2D motion model, the 2D feature points with missing depth information are added, which makes full use of the image information and improves the matching accuracy. Secondly, the map information of the key frame and the previous frame is considered to estimate iteratively, which increases the number of matching points and provides more constraint information. Finally, on the basis of the hybrid estimation method, the SBA method is combined to optimize the pose estimation results, which achieves high localization accuracy and small accumulation error. In this paper, experiments are performed on a Kinect-based mobile platform. The comprehensive offline datasets and online experimental tests show that this method not only meets the requirement of real-time, but also improves the accuracy of the mobile locali-zation effectively.

Key words: visual vdometry, Oriented FAST and Rotated BRIEF（ORB）, hybrid motion estimation, Sparse Bundle Adjustment（SBA）, Kinect