Correcting Perspective Distortion In Incremental Video Stitching

Abstract

In video stitching, only incremental video stitching technology can typically be used due to the limitations of computer memory and processing speeds. However, if frames with perspective distortion are selected as reference planes, perspective distortion will accumulate continuously during incremental video stitching, making the resulting panorama image appear unnatural. Existing solutions to correct perspective distortion rely on the content of scenes or external hardware, which lack generalization and yield poor performance when stitching UAV (unmanned aerial vehicle) videos. This paper proposes a novel type of feature called “displacements of matching feature points” to correct perspective distortion. A displacement vector composed of matching feature points on two adjacent frames contains the information of perspective distortion, which allows us to correct any perspective distortion. We correct the displacements of matching feature points with perspective distortion in a reference plane to one without perspective distortion. Then, this plane is taken as a global reference plane in incremental video stitching to further correct the panorama. Using the proposed method, correction of perspective distortion is no longer dependent on image content or computational hardware. Experimentally, the proposed method performs better than other state-of-the-art methods, yielding more natural panoramas.