Improving Visual Odometry by Removing Outliers in Optic Flow
Pedro Santana, Luís L. Correia · 2008
Abstract — Stereo based visual odometry is particularly interesting for affordable all-terrain robots, which would otherwise require expensive inertial navigation systems. A key aspect of the method is the process of matching image features across frames, which tends to produce a considerable amount of outliers, i.e. mismatched features. Outliers are usually handled in the subsequent motion estimation step, by recurring to robust statistical methods (e.g. RANSAC), whose computational cost grows with the probable number of outliers. Other approaches remove the outliers prior to the motion estimation step, guaranteeing a fixed computational cost independent of the number of outliers. This paper departs from previous work on this latter approach by proposing a heuristic algorithm which does not need to generate rotation hypotheses, nor to assume that the robot is moving on a planar surface. This algorithm is part of a wider real-time visual odometry system implemented on a stereo-based all-terrain robot. Experimental results in the physical robot highlight the capabilities of the proposed method. Index Terms — Visual odometry, outliers removal, stereo vision, optic flow, mobile robots.