A Method of Human Action Recognition Based on Spatio-Temporal Interest Points and PLSA

Human action recognition is a significant task in automatic understanding systems for video surveillance. A detector of spatio-temporal interest points is presented in this paper combining Gaussian Mixture Model (GMM) with 3D Hessian detector. The Features descriptors of 3DHOG/HOF are computed on the pixels around these interest points. We perform a unified action recognition framework based probabilistic latent semantic analysis (PLSA). Experiments on the CASIA dataset apply the leave-one-out and compare our proposed approach with other methods. The result shows that our proposed approach is effective. Compared with other approaches our approach is more robust, easier to compute.