Detecting lines and building intersection correspondences by computing edge oriented histogram on multi-sensor images

This paper proposes an approach to establishing correspondences for line intersections on multi-sensor images containing lines. Keypoints have been widely applied in a variety of computer vision fields. On multi-sensor images the number of keypoints may be much smaller than on single-sensor images due to the lack of texture. In addition, the multimodality often causes the incorrect assignment of main orientation to keypoints. Observing this, this paper proposes extracting line intersections as keypoints and utilizing one line as the main orientation to compute edge oriented histogram (EOH) descriptor. For EOH descriptor, gradient orientation is employed to compute the filter responses. Experimental results show that the proposed method can build more reliable keypoint matchings on challenging image pairs such as visible image and middle/long-wave infrared images.