Effects of different camera motions on the error in estimates of epipolar geometry between two dimensional images in order to provide a framework for solutions to vision based Simultaneous Localization and Mapping (SLAM)

Abstract

This thesis explores the effect camera motion and feature tracking have on the estimations of an epipolar geometry at different stages of a 3D reconstruction and relates the findings to a framework for vision based Simultaneous Localization and Mapping (SLAM). Although there have been previous attempts to determine the quality of algorithms that calculate a fundamental matrix, both robust and linear, we have found no study that explores the relationship between camera motion, or likewise the different types of parallax, and errors in the epipolar geometry between two images as defined by an estimated fundamental matrix. The interest comes from the fact that there are claims to this end made by two prominent textbooks in this area. By using synthetic scenes that are projected with and without noise by camera matrices that define different camera motions between the projections we are able to isolate the three different type of parallax that can be experienced between projections; no parallax shift from rotational movement, a high amount of parallax shift from translational movement in the camera's xy-plane, a high amount of parallax shift from translational movement along the camera's optical axis (z-plane). We also studied an unconstrained movement with components of each of the previous three types. The different camera motions are equivalent to different motions a robot would experience when performing SLAM, specifically, rotational, lateral, forward and unconstrained motions. There are multiple experiments that explore the effect motion has at every stage of a projective reconstruction algorithm.