Action Recognition Using Topic Models

In this book chapter, we will introduce approaches of using topic models for action recognition. Topic models were originally developed in language processing. In recent years, they were applied to action recognition and other computer vision problems, and achieved great success. Topic models are unsupervised. The models of actions are learned through exploring the co-occurrence of visual features without manually labeled training examples. This is important when there are a large number of actions to be recognized in a large variety of scenes. Most topic models are hierarchical Bayesian models and they jointly model simple actions and complicated actions at different hierarchical levels. Various knowledge and contextual information can be well integrated into topic models as priors. We will explain how topic models can be used in different ways for action recognition in different scenarios. For examples, the scenes may be sparse or crowded. There may be a single camera view or multiple camera views. The camera settings may be near-field or far-field. In different scenarios, different features, such as trajectories, local motions and spatial-temporal interest points, are used for action recognition.