Multiple Homography Estimation via Stereo Line Matching for Textureless Indoor Scenes

Abstract

For robot navigation applications, indoor scenes usually contain texture-less planar surfaces, which renders traditional point-based matching approaches deficient. As an emerging alternative, line-based methods have attracted growing interest. In our paper, we develop a unified approach for line detection and stereo matching that can produce appealing results in un-textured scenes. In addition, we propose a co-planar line classification approach, based on which we are able to enforce the epipolar line constraints to multiple homography compatibility validation. Experimental results demonstrate the efficiency of our method as well as its improved capability on multi-homography estimation over traditional methods.