A generative traversability model for monocular robot self-guidance

Abstract

In order for robots to be integrated into human active spaces and perform useful tasks, they must be capable of discriminating between traversable surfaces and obstacle regions in their surrounding environment. In this work, a principled semi-supervised (EM) framework is presented for the detection of traversable image regions for use on a low-cost monocular mobile robot. We propose a novel generative model for the occurrence of traversability cues, which are a measure of dissimilarity between safe-window and image superpixel features. Our classification results on both indoor and outdoor images sequences demonstrate its generality and adaptability to multiple environments through the online learning of an exponential mixture model. We show that this appearance-based vision framework is robust and can quickly and accurately estimate the probabilistic traversability of an image using no temporal information. Moreover, the reduction in safe-window size as compared to the state-of-the-art enables a self-guided monocular robot to roam in closer proximity of obstacles.