Region Classification for Robust Floor Detection in Indoor Environments

Abstract

A novel framework based on stereo homography is proposed for robust floor/obstacle detection, capable of producing dense results. Floor surfaces and floor anomalies are identified at the pixel level using the symmetric transfer distance from the ground homography. Pixel-wise results are used as seed measurements for higher lever classification, where image regions with similar visual properties are processed and classified together. Without requiring any prior training, the method incrementally learns appearance models for the floor surfaces and obstacles in the environment, and uses the models to disambiguate regions where the homography-based classifier cannot provide a confident response. Several experiments on an indoor database of stereo images with ground truth data validate the robustness of our proposed technique.