A Robust and Efficient Homography Based Approach for Ground Plane Detection

This paper presents a homography based ground planedetection method. The method is developed as a part of stereovision based obstacle detection technique for the visuallyimpaired people. The method assumes the presence of a texturedominant ground plane in the lower portion of the scene, whichis not severe restriction in a real world. SIFT algorithm is usedto extract features in the stereo images. The extracted SIFTfeatures are robustly matched by model fitting using RANSAC.A sample of putative matches lying in the lower portion of theimage is selected. A fitness function is developed to selectmatches from this sample, which are used to estimate groundplane homography hypothesis. The ground plane homographyhypothesis is used to classify the SIFT features as eitherbelonging to ground plane or not. Image segmentation usingmean shift and normalized cut is further used to filter theoutliers and augment the ground plane. Experimental testshave been conducted to test the performance of the proposedapproach. The tests indicate that the proposed approach hasgood classification rate and have operating distance rangefrom 3 feet to 12 feet.