Automated Place Detection Based on Coherent Segments

This paper is concerned with automated appearance-based place detection. This requires the robot to partition the incoming visual stream in a place-identity independent manner and determine the set of appearances belonging to each place. While previous work uses low-level features for place detection, we introduce a novel approach based on spatio-temporally coherent segmented regions. Our motivation is that these segments persist in the scene under a wider range of viewpoints and dynamical changes - based on our previous work on video summarization. In this approach, each incoming appearance is represented by a region adjacency graph that encodes the prevailing segments and their spatial relations. The robot detects a place through tracking the region adjacency graphs across the incoming appearance data based on the spatio-temporal coherency of their constituent segments. As such, place detection can be done more reliably while simultaneously generating a segment-based place representation that can be used in the ensuing semantic analysis of the place. This is demonstrated by a series of on-robot experiments using indoor and outdoor benchmark datasets including a comparative study.