Human action recognition via compressive-sensing-based dimensionality reduction

We propose a new dimensionality reduction method called compressive sensing with Gaussian mixture random matrix (CS-GMRM), in which a novel measurement matrix using Gaussian mixture distribution is constructed and is proved to satisfy the restricted isometry property. The CS-GMRM method projects high-dimensional vector spaces into low-dimensional ones via a single matrix multiplication. In particular, the proposed method removes the need of a training process, preserves the metric information of the original vector space, and requires a low level of computational complexity. We apply our method to the problem of recognizing human action from video sequences. Experimental results show that the proposed method is simultaneously highly effective and highly efficient for action recognition, and outperforms the state-of-the-art dimensionality reduction methods.