Towards fully self-supervised free space estimation for unmanned ground vehicles -

Abstract

With the era of autonomous robots about us, the problem of scene understanding has become particularly important. The transformation of robots from human super- vised systems to fully autonomous agents requires these robots to have a reliable un- derstanding of their environment. In its most basic form, this understanding reduces to delineating occupied space from free space. Such reliable detection of free space is essential for a system to safely navigate its environment and therefore is a major interest for the robotics research community. This thesis provides a system that automatically learns, classifies and maps free space in an environment using a stereo sensor. This is done through employing self- supervision by designing an algorithm that extracts training data automatically and reliably from free space in stereo data. The proposed algorithm is shown to be superior to other algorithms in literature by benchmarking on three stereo datasets. The results of free space classification and mapping are presented and analyzed to further validate the proposed system.