View-invariant action recognition using interest points

Abstract

In this paper, we present a two-layer classification model for view-invariant human action recognition based on interest points. Training videos of every action are recorded from multiple viewpoints and represented as space-time interest points. These videos do not require temporal aligning and camera estimating. The first layer of the model is view clustering. We cluster all the videos of an action using K-Means, and break the action into several sub-actions. The second layer is Bayes classifying. We use Naive Bayes to train the sub-classifiers for the sub-actions, and then generate an optimal classifier for the action. Unlabeled data can be recognized by the optimal classifiers, which may be single-view videos, multi-view videos, or long multi-action videos. Finally, we test our algorithm on the IXMAS dataset, and the CMU motion capture library. The experiments demonstrate that our algorithm can recognize the view-invariant actions and achieve high recognition rates.