Human action recognition using topic model

In this paper, we propose a novel method to recognize actions by using latent topic model which can be used for fusion of multi-view HOG3D feature and location and recognition of local space-time interest point. The latent topic model describes the probability distribution of spatial-temporal words and the intermediate topics about the action categories so that we can tag each spatial-temporal word with action categories. That means we can easily localize and recognize multiple actions. In our method, we first extract the spatial-temporal interest point and the multi-view HOG3D feature of the cuboid around the interest point. Then we construct the vocabulary and utilize the latent topic model to learn the intermediate topics. The proposed approach is tested on publicly available MSRAction3D dataset, revealing the advantages and the state-of-art performance of our method.