Analysis of Human Actions for Video Indexing

摘要

Automatically understanding human actions is crutial for efficiently indexing many types of videos, such as sports videos, home videos, movies etc. However, it is challenging due to their variances caused by different actors, different scales, and different views. In order to incorporate these variances, most methods in literature have to sacrifice the discriminability of action models. In this paper, we address the tradeoff between invariability and discriminability. We firstly propose a novel set of pixel-wise features which are invariant to actor appearances, scales, and motion directions. Then, multi-prototype action models are constructed to realize view invariance. By leaving the most challenging invariance from feature level to model level, we successfully maintain the discriminability of action models. The extensive experiments demonstrated the good performance of the proposed method.