Subspace learning for silhouette based human action recognition

This paper exploits different subspace learning methods applied on silhouette based action recognition and evaluates their performance. Our recognition scheme is formed by segmenting action sequence into overlapped sub-clips and using sub-models for action matching. This sub-model matching method shows advantages in processing periodic actions. The experimental results prove that human action silhouettes are very informative for action recognition and subspace analysis can effectively preserve the intrinsic structure of raw data from 3D silhouettes. The subspace learning methods compared in this paper include traditional methods - Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA), and recently reported Orthogonal Local Preserving Projection (OLPP). PCA is observed to perform the best regarding both accuracy and efficiency. We believe our work is helpful for further research in silhouette based action recognition combined with subspace learning methods.