Abstract: Human action recognition from video sequences is a challenging problem in computer vision. This paper uses the bag-of-words paradigm inherited from text analysis to represent a clip of video as a document. The local features are extracted from spatio-temporal interest points which are points with local maximum variation in both space and time domain. Then topic models on video documents are applied to analyze the latent topics and actions in the video are recognized in a discriminative fashion. The proposed system is tested on both simple and complex data sets. Experiment result shows that the approach is comparable or better than all published state-of-the-art methods.

Key words: human action recognition, spatio-temporal interest point, bag-of-words model, topic model