Spatial Uncertainty In Stereo Visual Odometry

It is worthwhile of the evaluation how VO performs by its uncertainty, and also appealing to improve its performance with knowledge of uncertainty. VO involves many general and core algorithms in the area of computer vision. This paper focuses on these key stages in stereo visual odometry, and performs rigorous mathematical analysis on its uncertainty. Optimization based on feature uncertainty is also studied to improve the accuracy of VO.1. This paper first introduces basic theory of stereo vision, and the framework of VO, including feature detection and matching, and RANdom SAmple Consensus which is used to remove outliers.2. On the uncertainty of image features, a framework of scale-invariant detectors is summarized. A general method for uncertainty estimation is developed with it. Simulation with SIFT detector proves the correctness of the formulae. Results are also compared with existing method.3. Uncertainty of motion parameters is studied. Uncertainty is propagated from image features to 3D points, and then to motion parameters. A linear method and a non-linear one are developed. Considering that RANSAC is an important part in VO, ignoring it may suffer from over-estimated uncertainty. Here the author adopts the method of Bayesian Inference to deal with the non-differentiable system with large amount of input points. By accumulating motions step by step, the uncertainty is also accumulated to the current location and orientation of robot.4. Finally, this paper adopts a method of maximum-likelihood estimation which concerns the uncertainty of image features to improve the accuracy of single-step motion estimation. Then an algorithm of local binocular bundle adjustment is proposed to further refine the estimated results. Series of experiments are conducted in various kinds of environments, including simulated lunar surface, outdoor lawn, and large-scale campus environment. Results are examined with the ground truth obtained by a high-precision total station, which verifies that our methods improve the accuracy of the system...