Robust Dimensionality Reduction for Human Action Recognition

Human action recognition can be approached by combining an action-discriminative feature set with a classifier. However, the dimensionality of typical feature sets joint with that of the time dimension often leads to a curse-of-dimensionality situation. Moreover, the measurement of the feature set is subject to sometime severe errors. This paper presents an approach to human action recognition based on robust dimensionality reduction. The observation probabilities of hidden Markov models (HMM) are modelled by mixtures of probabilistic principal components analyzers and mixtures of $t$-distribution sub-spaces, and compared with conventional Gaussian mixture models. Experimental results on two datasets show that dimensionality reduction helps improve the classification accuracy and that the heavier-tailed $t$-distribution can help reduce the impact of outliers generated by segmentation errors.