Visual Odometry based on improved feature matching and Unscented Kalman Filter

In this paper, we present an improved vision-based navigation method and proposed an improved feature matching method for improving the matching accuracy. In the matching process, we divide it into two steps, coarse and fine matching. During the coarse matching step, we adopt SURF feature detector for feature detection and Fast Library for Approximate Nearest Neighbors for feature matching, and then use the constraints of epipolar geometry, major orientation of feature points, and the uniqueness of feature matching to roughly eliminate error matching. In the fine matching process, Random Sample Consensus method with outlier rejection is employed, which will reduce the effects on motion estimation by moving objects in the scenes. The visual odometry algorithm is based on trifocal geometry, which is no need for the reconstruction of the 3d object points. Finally, we employ Unscented Kalman Filter for ego-motion estimation, which is better than Extended Kalman Filter and the experimental result shown that it can fully adapt to environment with high uncertainty. The experimental results prove that the method proposed in this paper is superior to other algorithm in terms of positioning precision.