D. K. Ufuktepe, J. Collins, H. AliAkbarpour, P. Narayanan, G. Seetharaman, and K. Palaniappan

Geospatial Informatics X, Volume 11398, pgs. 113980B, 2020

PlainText, Bibtex, PDF, Google Scholar

Abstract

Fast, efficient and robust algorithms are needed for real-time visual tracking that could run smoothly onboard airborne embedded systems. In order to efficiently use a motion-detection-based tracking algorithm on a raw image sequence captured by a moving platform, the relative motion between the camera and the ground plane must be stabilized first so that only the real moving objects in the scene could be detected. Using point correspondences to estimate the homography matrix between the frames is a common registration method used for the stabilization. In order to have a valid homography estimation, most of the point correspondences between the images have to lie on the same 3D plane (i.e. ground plane). However, when the scene includes multiple complex 3D structures (e.g. trees, cloudy sky, buildings) it becomes extremely challenging to satisfy this constraint, particularly for aerial images acquired from a low altitude flying drone with an oblique camera angle imaging a scene in which the ground plane is not the dominant object.