Temporal texture characterization for block-based video indexing

Temporal texture accounts for a large proportion of motion commonly experienced in the visual world. Contemporary video indexing methods classify temporal texture videos either using some spatial cooccurrence matrices on a limited number of representative or merged frames or using temporal cooccurrence matrices over all the frames in the sequence. However, classifying temporal textures using either space or time domain information, while ignoring the other domain, has serious flaws. In this paper, a new video indexing method is developed to classify temporal texture videos using spatiotemporal cooccurrence matrices over the entire frame sequence thus utilizing both domains. Although dynamic content characterization is generally performed on the analysis of pixel-based optical flows or normal flows or normalised motion related measures, the proposed method is developed to classify block-based videos in order to take the advantage of using already available motion vectors to keep the computational complexity within realtime bound. Experimental results clearly demonstrate the ability of the proposed method to characterize temporal textures in both motion magnitude and direction domains.