Human Action Recognition Using Manifold Learning and Hidden Conditional Random Fields

A model-based probabilistic method of human action recognition is presented in this paper. We employ supervised neighborhood preserving embedding (NPE) to preserve the underlying structure of articulated action space during dimensionality reduction. Generative recognition structures like Hidden Markov Models often have to make unrealistic assumptions on the conditional independence and can not accommodate long term contextual dependencies. Moreover, generative models usually require a considerable number of observations for certain gesture classes and may not uncover the distinctive configuration that sets one gesture class uniquely against others. In this work, we adopt hidden conditional random fields (HCRF) to model and classify actions in a discriminative formulation. Experiments on a recent database have demonstrated that our approach can recognize human actions accurately with temporal, intra- and inter-person variations.