Toward Lifelong Object Segmentation from Change Detection in Dense RGB-D Maps

Abstract

In this paper, we present a system for automatically learning segmentations of objects given changes in dense RGB-D maps over the lifetime of a robot. Using recent advances in RGB-D mapping to construct multiple dense maps, we detect changes between mapped regions from multiple traverses by performing a 3-D difference of the scenes. Our method takes advantage of the free space seen in each map to account for variability in how the maps were created. The resulting changes from the 3-D difference are our discovered objects, which are then used to train multiple segmentation algorithms in the original map. The final objects can then be matched in other maps given their corresponding features and learned segmentation method. If the same object is discovered multiple times in different contexts, the features and segmentation method are refined, incorporating all instances to better learn objects over time. We verify our approach with multiple objects in numerous and varying maps.