This paper presents a probabilistic method of human action recognition based on manifold learning and Hidden Conditional Random Fields (HCRF). A supervised Neighborhood Preserving Embedding (NPE) is employed for dimensionality reduction by preserving the local neighborhood structure on the data manifold. Most existing approaches to action recognition use a Hidden Markov Model or suitable variant to model actions; a significant limitation of these models is the requirements of conditional independence of observations. In addition, generative models are selected to maximize the likelihood of generating all the examples of a given class and may not uncover the distinctive configuration that sets one class uniquely against others. HCRF relaxes the independence assumption and classifies actions in a discriminative hidden-state formulation. Experimental results on a recent database have demonstrated that this approach can recognize human actions accurately with temporal, intra- and inter-person variations even when noise and other factors such as partial occlusion exist.