Manifold-constrained coding and sparse representation for human action recognition

Due to its various applications, human action recognition has been widely studied and achieved tremendous progress. However, how to learn an accurate and discriminative behavior representation based on the extracted features remains as a challenging problem. In this paper, we present an effective coding scheme that can discover the manifold structure of the learned features with an l2-norm regularization. Coupled with a local constraint, the proposed coding scheme, which has an analytical solution can learn an accurate, compact and yet discriminative behavior representation. After the behavior representations are obtained, the action recognition problem is formulated as a sparse linear representation of an overcomplete dictionary constructed by labeled behavior representations. The same manifold l2-norm regularization is also employed in this stage. The reconstruction error associated with each class is used for classification. Experimental results demonstrate the effectiveness of the proposed approach on several public datasets including various physical actions and facial expressions.