Improving video-based robot self localization through outlier removal

The purpose of this paper is to present a method for rejecting false matches of points from successive views in a video sequence - e.g., one used to perform Pose from Motion for a mobile sensing platform. Invariably, the algorithms used to determine point correspondences between two images output false matches along with the true. These false matches negatively impact the calculations required to perform the pose estimation from video. This paper presents a new algorithm for identifying these false matches and removing them from consideration in order to improve system performance. Experimental results show that our algorithm works in cases where the percentage of false matches may be as high as 80%, providing a set of point correspondences whose true/false match ratio is much higher than the mutual best match method commonly used for outlier filtering, resulting in comparable or better outlier rejection - increasing the true/false match ratio by 2-3 times - in only a fraction of the time. Robot self-localization,.