Image registration for video stabilization and motion detection

A hybrid image registration method for the purpose of stabilizing a video captured by a moving camera is presented. The following assumptions are made: 1) The provided video contains moving targets. 2) The camera capturing the video is sufficiently far from the scene so that variation in scene depth across an image is negligible compared to the distance of the camera to the scene. 3) The area covered by the moving targets in an image is only a small portion of the image. Various intensity-based and featurebased registration methods are evaluated and use of geometric constraints to register and stabilize videos is examined. Experimental results show that a hybrid method that uses a combination of landmarks, intensity-based matching, and geometric constraints best registers and stabilizes a video captured by a moving camera.