Improving feature based object recognition in service robotics by disparity map based segmentation

This article presents a novel object recognition module which is adapted to the needs in mobile service robotics. It uses information provided from a stereo camera system as pre-processing part of SIFT or SURF. The principle idea is to filter irrelevant information by selecting regions of interest in the disparity map from stereo images and to use the geometrical constraints of the stereo camera system in order to filter out useless descriptors in early stages of the processing chain. Experimental results show that this setup improves overall performance in comparison to similar systems by a factor of two to five.