Visual Odometry Based on 3D-3D and 3D-2D Motion Estimation Method

The depth information of RGB-D camera is limited and susceptible to noise, in order to improve the accuracy of visual odometry, a modified motion estimation method based on 3D-3D and 3D-2D is proposed. Firstly, the ORB feature points of the current frame image are extracted and feature matching is performed with adjacent frames to achieve a higher matching accuracy. Secondly, using a 3D-3D model based on the RICP algorithm (RANSAC-ICP), combined with the 3D-2D motion model considering keyframes, and iteratively estimates the interframe motion, thereby increasing the number of matching pairs and providing more constraint information. Finally, local optimization is performed by the Sparse Bundle Adjustment method(SBA) to optimize estimated pose, so that the camera pose and map point can reach the minimum error at the same time. The experiments are performed on mobile platform with Kinect camera. The experimental tests show that this method not only meets the requirement of real-time location, but also effectively improves the accuracy of the localization.