Spatio-Temporal Video Segmentation

This Chapter considers approaches to video skimming into semantically-consistent segments of video streams, which are highly redundant and weakly structured data. In such a way, one of the promising ways is spatial-temporal segmentation as frame partitions represent certain spatial image content. Also, properties were formulated and proved which ultimately determine the characteristics of permissible segmentation transformations when searching for a compromise between over and undo segmentation. Temporal segmentation of multidimensional time series has been examined, which enables structuring video streams and significantly reduce the amount of data that will require online processing. For this, multidimensional time series analysis theory was used, since a completely natural video representation is a sequence of frames, followed by their combination into groups of frames (shots) with the same content. It was shown that various approaches can be used to detect shots with homogeneous characteristics, which are based on VAR models, exponential smoothing and predictive models.