Image Registration Techniques Based on the Scale Invariant Feature Transform for Aerial Surveillance Images

Aerial surveillance plays a vital role in continuous monitoring of areas where human intervention is of higher risk. Image registration is a process to bring several images of a scene to a single co-ordinate system. It is widely done using the local features extracted from the images. Local features of an aerial surveillance image is also used in several computer vision tasks such as scene matching, object identification and object tracking. The features identified must be repeatable across several transformations on the image.？ This paper aims to find the most repeatable local feature from the existing state of art affine invariant features : Harris Affine, Hessian Affine. They are described using the Scale Invariant Feature Transform (SIFT). The dataset comprises aerial images captured from low altitude unmanned aerial vehicles using thermal and visual sensors. The ground truth between the images of a scene is estimated manually using control points, in the form of a homograph matrix. The repeatability of the various features are estimated using the ground truth estimated. The Hessian Affine feature outperforms in performance compared to the other affine invariant feature detector making it a potential candidate for the various computer vision tasks on aerial surveillance images