Mapping through a stereo rig : identifying ground patches and obstacles -

Abstract

Mapping the space about an autonomous agent is one of the preliminary steps for any sort of navigation or task planning. In this work, stereo rig output is employed to accomplish this task, where ground patches, near ground obstacles and background are delineated in one of the stereo images, in a self-supervised method. In a first method, free space is estimated using a novel technique: first depth information extracted from a stereo camera is used to identify, with high certainty, regions that belong to ground or obstacle clutter. Next, interactive graph cuts is used to propagate the initial segmentation across the entire image to yield an estimate of the location of free space in each image, while considering all obstacles and background as one object. This novel method is proved to outperform state of the art in the literature; however, to mitigate the effect of local noise in the disparity image, and reduce the runtime of the system, the self supervised classifier is used to train a ground segmentation deep learning algorithm, which is fine-tuned online when the system is faced with new ground. The system detects the change in ground using a novel image clustering technique, and by comparing its output against a weak classifier.