Fast Non-Linear Video Synopsis

Abstract

This paper presents a very efficient method for extracting the non-linear summary of a video sequence by synthesizing new summary frames from a number of original frames. The non-linear summary of a video sequence is fundamentally different from the classical video summarizing techniques which discard full frames. The high efficiency of the method is due to the employment of dynamic programming to find the space-time surfaces in the video volume that have lesser motion information. We project the video volume onto a plane orthogonal to one of its axes and find a minimum cost path on the time derivative of the projected image. Back-projecting the minimum cost path down to the video volume gives the space time surface to be discarded. Applying this process several times results in the summarized video. One of the major contributions of our method is that it can work on real-time video sequences with a small memory requirement. We demonstrated our method on several examples which are available at http://vision.gyte.edu.tr/projects.php?id=5.