CNN-LSTM-Network Based Robot Localization Using Spherical Images of Neighboring Places

Abstract

The weakness of vision-based localization methods is that a robot may be confused when similar scenes are observed. In this paper we propose a method to cope with this problem by using multiple spherical images captured at neighboring places; that is, the ambiguity caused by visual similarity at a place is decreased by observing the scenes of its neighboring places. Concretely, the route that a robot moves along is sampled by discrete points which correspond to the places to be localized. Then, the localization problem is formulated as a classification problem, in which a place (a sampling point along a route) corresponds to a class. A CNN Long-Short-Term-Memory network is implemented to realize this task. The effectiveness of the proposed method is shown in the preliminary experiment.