Spatiotemporal Oriented Energy Based Dynamic Texture Research

Dynamic texture is an important problem in computer vision researchcommunity. It comprises a huge class of visual motion patterns of spatialrepetitiveness and temporally stationarity. Research on dynamic texturerepresentation, recognition, segmentation and other applications helpenrich our knowledge of complex visual motion understanding incontent-based video retrieval and motion recognition among many others.We base our research on the nonparametric motion representation ofdistributed spatiotemporal oriented energy that characterizes appearanceand dynamics jointly. This paper presents work in the following respects.First, we present an integration of spatiotemporal oriented energy andlocal binary pattern (LBP) to model local dynamic texture. It extractssufficient information of coarse motion clues and local structure statistics.Multiple weighted histograms of the resulting local spatiotemporaloriented energy binary patterns are responsible for representing andmeasuring regional motion. Based on this formulation, an efficient andimproved Split-merge segmentation method is implemented. Extensiveexperimental results on both synthesized and natural scenes verify that theproposed method performs well with a low computational complexity.This paper also presents a computational unsupervised saliency modelutilizing bio-inspired dynamic texture feature for spatiotemporal saliency.We first propose distributed opponent oriented energy for capturing localdynamic texture ensemble motion. Then, we generalize dynamic textureself-resemblance for spatiotemporal saliency. High effectiveness andefficiency of the proposed method is demonstrated both qualitatively and quantitatively, for background subtraction in the cases of extremelydynamic scenes and camera jitter. In terms of the trade-off betweenaccuracy and computation cost, our method achieves competitive results incontrast to the state-of-art algorithm.