Feature correspondence finding with vertical cylinder and epipolar geometry for indoor environments

We introduce an improved approach, called FCVC, that finds accurate matches from two related images of indoor environments by matching corresponding vertical cylinders of features, then retrieves more matches by epipolar geometry. Feature matching is a critical technique for many vision-based applications. However, since the preliminary matching results obtained by SIFT sometimes are not accurate enough, and therefore, post processing methods, like RANSAC, have been proposed to increase the correctness of feature mappings. As RANSAC iteratively selects partial matches from preliminary matches and recovers the epipolar geometry to find correct matches, our method first finds accurate matches in vertical cylinders, based on three consistency properties which are common while two images are taken on a flat floor: (1) altitudinal order, (2) composition of features for each vertical cylinder, and (3) horizontal order of all vertical cylinders. Once some accurate matches are acquired, the epipolar geometry is recovered to retrieve matches that are not grouped into any vertical cylinder previously. Experiments in different indoor environments reveal that the proposed FCVC in general spends at most 6% of the time required by RANSAC, has at least 71% of the number of matches obtained by RANSAC, and yet with accuracy rate similar to that of RANSAC.