Human Action Recognition Based on Discriminative Sparse Representation on Multi-Manifolds

Human action recognition is an important problem in computer vision. One of the methods that are recently used is sparse coding. Conventional sparse coding algorithms learn dictionaries and codes in an unsupervised manner and neglect class information that is available in the training set. But in this paper for solving this problem, we use a discriminative sparse code based on multi-manifolds. We divide labeled data samples into multi-manifolds and also to decrease run time, reduce dimension of manifolds. We find k inter nearest neighbors and intra nearest neighbors for each data sample in each manifold. The intra class variance should be minimized while the inter class variance should be maximized, in the result we could calculate laplacian matrix and optimize sparse code and dictionary. Then we use discriminative sparse error for classification. We run this method on KTH and UCF sport datasets. Results show that we obtain a better result (about 89%) in UCF dataset.