Nature-inspired, parallel object recognition

Abstract

In this article we present a massively parallel object recognition system designed to operate on-line, processing data acquired by indoor mobile robots equipped with 3D cameras. Inspired by the properties of the mammalian visual cortex, the proposed method incorporates a learned, selective use of features for the recognition of specific objects of interest, as well as a pre-processing stage of simultaneous localization and mapping (featuring Kinect Fusion) and a new parallel, heuristic scene segmentation algorithm. The benefits of applying class-specific feature spaces are demonstrated in an experiment carried using indoor scenes containing multiple common household objects.