The stitching of aerial videos from UAVs

Image stitching has been extensively studied over the last decade, particularly since the introduction of scale invariant feature transform and bundle adjustment techniques. This paper investigates three major challenges that need to be solved for aerial video stitching: significant motion blur in some frames, the scale factor in the stitched images and the constant change of the camera projection centre. We propose to estimate the degree of blurring for each video frame and to avoid using frames with significant motion blur if possible. We propose to detect objects with obvious depth changes and to select a region with such objects for stitching. We also proposed novel normalisation approach for scale correction. As a result, we are able to obtain quality stitched images covering large ranges.