Wide-area egomotion from omnidirectional video and coarse 3D structure

Abstract

This thesis describes a method for real-time vision-based localization in human-made environments. Given a coarse model of the structure (walls, floors, ceilings, doors and windows) and a video sequence, the system computes the camera pose (translation and rotation) in model coordinates with an accuracy of a few centimeters in translation and a few degrees in rotation. The system has several novel aspects: it performs 6-DOF localization; it handles visually cluttered and dynamic environments; it scales well over regions extending through several buildings; and it runs over several hours without losing lock. We demonstrate that the localization problem can be split into two distinct problems: an initialization phase and a maintenance phase. In the initialization phase, the system determines the camera pose with no other information than a search region provided by the user (building, floor, area, room). This step is computationally intensive and is run only once, at startup. We present a probabilistic method to address the initialization problem using a RANSAC framework. In the maintenance phase, the system keeps track of the camera pose from frame to frame without any user interaction.

(cont.) This phase is computationally light-weight to allow a high processing frame rate and is coupled with a feedback loop that helps reacquire "lock" when lock has been lost. We demonstrate a simple, robust geometric tracking algorithm based on correspondences between 3D model lines and 2D image edges. We present navigation results on several real datasets across the MIT campus with cluttered, dynamic environments. The first dataset consists of a five-minute robotic exploration across the Robotics, Vision and Sensor Network Lab. The second dataset consists of a two-minute hand-held, 3D motion in the same lab space. The third dataset consists of a 26-minute exploration across MIT buildings 26 and 36.