Action Recognition Using Topic Models

Abstract

In this book chapter, we will introduce approaches of using topicmodels for action recognition. Topic models were originally developedin language processing. In recent years, they were applied to actionrecognition and other computer vision problems, and achieved greatsuccess. Topic models are unsupervised. The models of actions arelearned through exploring the co-occurrence of visual features withoutmanually labeled training examples. This is important when thereare a large number of actions to be recognized in a large varietyof scenes. Most topic models are hierarchical Bayesian models andthey jointly model simple actions and complicated actions at differenthierarchical levels. Knowledge and contextual information can bewell integrated into topic models as priors. We will explain howtopic models can be used in different ways for action recognitionin different scenarios. For examples, the scenes may be sparse orcrowded. There may be a single camera view or multiple camera views.The camera settings may be near-field or far-field. In differentscenarios, different features, such as trajectories, local motionsand spatial-temporal interest points, are used for action recognition.