Human Action Recognition: An Innovative Approach Using Dynamic Affine-Invariants and Spatio-Temporal Action Features

摘要

In this paper, we introduce an innovative technique for recognizing human activity based on three main phases. The first phase of this approach is to an employ the modelling background that uses an adaptive Gaussian mixture to distinguish moving foregrounds from their moving cast shadows.Secondly, the extracted features of dynamic affine-invariants as Zernike moments and elliptic Fourier, in addition to mass centre and optical flow motion features are derived from 3D spatiotemporal action volume. Finally, the discriminative model of Hidden Conditional Random Fields (HCRFs)performs the training and testing action processes employing the combined features that execute powerful view-invariant process. Our experiment on benchmark action of KTH and Weizmann datasets, as well as UCF dataset demonstrate that the proposed approach is robust and efficient to the previousresearch. It also can proceed with no immolate real-time performance for an enormous area of workable action implementations.