Scene classification using unsupervised neural networks for mobile robot vision

Abstract

This paper presents an unsupervised scene classification method based on context of features for semantic recognition of indoor scenes used for an autonomous mobile robot. Our method creates Visual Words (VWs) of two types using Scale-Invariant Feature Transform (SIFT) and Gist. Using the combination of VWs, our method creates Bags of VWs (BoVWs) to vote to a two-dimensional histogram as context-based features. Moreover, our method generates labels as a candidate of categories with maintaining stability and plasticity together using the incremental learning function of Adaptive Resonance Theory-2 (ART-2). Our method actualizes unsupervised learning based scene classification using generated labels of ART-2 for teaching signals of Counter Propagation Networks (CPNs). The spatial and topological relations among scenes are mapped on the category map of CPNs. The relations of classified scenes that contain categories are visualized on the category map. The experiment demonstrates that classification accuracy of semantic categories such as office rooms, corridors, etc. using an open dataset for an evaluation platform of position estimation and navigation for an autonomous mobile robot.