Action recognition in depth videos using nonparametric probabilistic graphical models

Action recognition involves automatically labelling videos that contain human motion with action classes. It has applications in diverse areas such as smart surveillance, human computer interaction and content retrieval. The recent advent of depth sensing technology that produces depth image sequences has offered opportunities to solve the challenging action recognition problem. The depth images facilitate robust estimation of a human skeleton’s 3D joint positions and a high level action can be inferred from a sequence of these joint positions. A natural way to model a sequence of joint positions is to use a graphical model that describes probabilistic dependencies between the observed joint positions and some hidden state variables. A problem with these models is that the number of hidden states must be fixed a priori even though for many applications this number is not known in advance. This thesis proposes nonparametric variants of graphical models with the number of hidden states automatically inferred from data. The inference is performed in a full Bayesian setting by using the Dirichlet Process as a prior over the model’s infinite dimensional parameter space. This thesis describes three original constructions of nonparametric graphical models that are applied in the classification of actions in depth videos. Firstly, the action classes are represented by a Hidden Markov Model (HMM) with an unbounded number of hidden states. The formulation enables information sharing and discriminative learning of parameters. Secondly, a hierarchical HMM with an unbounded number of actions and poses is used to represent activities. The construction produces a simplified model for activity classification by using logistic regression to capture the relationship between action states and activity labels. Finally, the action classes are modelled by a Hidden Conditional Random Field (HCRF) with the number of intermediate hidden states learned from data. Tractable inference procedures based on Markov Chain Monte Carlo (MCMC) techniques are derived for all these constructions. Experiments with multiple benchmark datasets confirm the efficacy of the proposed approaches for action recognition