Dynamic texture recognition using sparse spatiotemporal derivative pattern

摘要

We present Spatiotemporal Derivative Pattern (SDP), a descriptor for dynamic textures. Using local continuous circular and spiral neighborhoods within video segments, SDP encodes the derivatives of the directional spatiotemporal patterns into a binary code. Moreover, an extension of SDP is introduced using the sparse coding method to represent the dynamic texture. The co-occurrence of two directional derivatives with same direction generates the dictionary matrix. Using the generated dictionary and the matching pursuit algorithm, the sparse representation of the dynamic texture is obtained. The main strength of SDP is that it uses fewer frames per segment to extract more distinctive features for efficient representation and accurate classification of the dynamic textures. The proposed SDP is tested on the Honda/UCSD and the You Tube face databases for video based face recognition and on the Dynamic Texture database for dynamic texture classification. Comparisons with existing state-of-the-art methods show that the proposed SDP achieves the overall best performance on all three databases. To the best of our knowledge, our algorithm achieves the highest results reported to date on the challenging YouTube face database.