Video-based Human Activity Analysis: An Operator-based Approach

Human activity in video sequences may be viewed as a sampled trajectory on a low dimensional manifold embedded in a high dimensional ambient space. Due to the unknown underlying manifold structure of the image frames, we propose a novel framework to define a neighborhood for high dimensional data which, when acted upon by a mapping operator, results in a subset in an a priorily well defined range space. We exploit the so-called correlation filtering with a specifically selected output response to effectively approximate the data manifold by way of encoding local neighborhoods on it. This helps us propose an unsupervised learning algorithm of human activity, and demonstrate its performance in classifying and clustering of different activities taking place in observed video sequences.