Perspective registration and multi-frame super-resolution of license plates in surveillance videos

One task often encountered in surveillance videos is the recognition of a target—e.g. the license plate of a vehicle. Often, the quality of a single video frame does not permit a reliable recognition. If multiple frames are available, it is possible to combine them in order to generate a single image with lower noise (frame averaging) and/or higher resolution (super-resolution). In order for these techniques to work, it is necessary to accurately estimate the motion of the object of interest in the recorded footage. In this paper, we introduce a method capable of accurately computing the perspective transformation that describes the motion of a planar object. The method minimizes the squared distance between the transformed image and a reference, computed over a user-defined region of interest, and uses the partial derivatives in order to significantly speed up the computation. This approach is inspired by the well known Kanade–Lucas–Tomasi feature tracker.