View-invariant action recognition based on temporal and spatial segmentation in 3D templates

Action representation is an important step in view-invariant action recognition. This paper introduces a novel representation based on the spatial segmentation and the extended temporal segmentation. First, representation based on special segmentation is introduced that is performed by separating moving part from action templates. Second, the Extended Representation Based on Temporal Segmentation (E-RBTS) is introduced. This method is created to improve Representation Based on Temporal Segmentation (RBTS). RBTS extracts subsequences of an action that contains important information about that action. Cylindrical Fourier transform is used in the proposed representation templates for view-invariant action recognition. Also the normalized correlation coefficient similarity measure is utilized in the classification step. This criterion is utilized for accurate discrimination between test feature and trains features. Experimental results on INRIA IXMAS dataset with 96.67% accuracy confirme the significant properties of new representation in the proposed method.