Representing and updating objects' identities in semantic SLAM

Abstract

Simultaneous localization and mapping (SLAM) deals with localizing and mapping in unknown environment. Semantic SLAM incorporates an additional layer of objects identities and their relationships. Here we suggest representing the identity of an object in semantic SLAM as a probability distribution over the object's traits, such as labels, colors, shapes, materials, etc. Objects' identities are estimated by integrating measurements from different sensors and are distinguished based on the discrepancy between the underlying probability distributions as quantified by the Bhattacharyya distance. The semantic mapping scheme is tested both in simulation and experiment using a ground robot.