Techniques for fast partitioning of compressed and uncompressed video

Video partitioning is the segmentation of a video sequence into visually independent partitions, which represent various identifiable scenes in the video. It is an important first step in considering other issues in video databases management, such as indexing and retrieval. As video partitioning is a computationally intensive process, effective management of digital video requires highly efficient techniques for the process. In general, for compressed and uncompressed video, the basic mechanism used to reduce computation is by selective processing of a subpart of the video frames. However, so far the choice of this proportion has been made randomly without any formal basis. An ad hoc selection of this subpart cannot always guarantee a reduction in computation while ensuring effective partitioning. This paper presents formal methods for determining the optimal window size and the minimum thresholds which ensure that decisions on scene similarity are made on a reliable, effective and principled basis. Further, we propose the use of neighbourhood-based colour ratios, and derive the ratio feature for both uncompressed and transform coded video. The neighbourhood-based ratio features account for both illumination variation and possible motion in the video, while avoiding the computational burden of explicit motion compensation procedures. Empirical results showing the performance of the proposed techniques are are also presented.