文摘
We propose a method for human action recognition using latent topic models. There are two main differences between our method and previous latent topic models used for recognition problems. First, our model is trained in a supervised way, and we propose a two-level Beta process hidden Markov model which automatically identifies latent topics of action in video sequences. Second, we use the human skeleton to refine the spatial–temporal interest points that are extracted from video sequences. Because latent topics are derived from these interest points, the refined interest points can improve the precision of action recognition. Experimental results using the publicly available “Weizmann”, “KTH”, “UCF sport action”, “Hollywood2”, and “HMDB51” datasets demonstrate that our method outperforms other state-of-the-art methods.