Histogram of Body Poses and Spectral Regression Discriminant Analysis for Human Action Categorization

This paper explores a recently proposed and rarely reported subspace learning method, Spectral Regression Discriminant Analysis (SRDA), on silhouette based human action recognition. The recognition algorithm adopts the Bag of Words (BoW) model combined with the action representation based on Histogram of Body Poses sampled from silhouettes in the video sequence. In addition, we compare the performance of SRDA for dimensionality reduction with several traditional subspace learning methods, such as Principle Component Analysis (PCA), supervised Locality Preserving Projections (LPP), unsupervised LPP and Neighbourhood Preserving Embedding (NPE). Experimental results show that Histogram of Human Poses combined with SRDA or its kernel version, SRKDA, can achieve 100% recognition accuracy for the Weizmann human action dataset, which is better than any published results on the same dataset.