Simultaneous Localization and Its Reliability Estimation using CNN and RBPF

Abstract

This paper presents a novel localization approach that simultaneously estimates a robot's pose and reliability of its estimation. To estimate the reliability, a convolutional neural network (CNN) is used as a decision maker for distinguishing whether localization has failed. The CNN, however, sometimes makes wrong decisions. To reduce influence of the wrong decisions, Rao-Blackwellized particle filter (RBPF) is employed. The reliability can be robustly estimated using the RBPF and it exactly describes successful and failure localization results. Exact performance of the reliability is shown through the experiments.