Road Plane Detection using Differential Homography Estimated by Pair Feature Matching of Local Regions

This paper presents a novel algorithm for road plane detection from an on-board camera. The algorithm employs the temporal difference of homography matrix, which is termed differential homography, caused by camera motion. Differential homography is estimated from optical flows of road plane regions, while using RANSAC algorithm to extract the majority optical flows. Since differential homography is estimated using the relationship between the image coordinate (location in an image) and the flows at the locations in an image. The proposed algorithm does not require the estimation of the homography matrix itself. Therefore, the proposed algorithm can be applied without calibration. The proposed algorithm effectively detect the optical flows from road region with using the pixel-pair feature matching. The algorithm is applied to the city traffic images distributed by UCL, and its average road detection ratio is found to be 75.6%. It is also applied to the previously collected suburban traffic images. A suitable detection result is obtained.