Non-negative sparse coding for human action recognition

We consider the problem of human action recognition using non-negative sparse representation of extracted features from spatiotemporal video patches. Our algorithm trains dictionaries for the calculation of a non-negative sparse representation for feature vectors and uses a linear Support Vector Machine (SVM) to distinguish between different actions. We evaluate the performance of the proposed techniques by using two human action datasets (KTH and IXMAS). In both cases, the proposed technique outperforms state-of-the-art techniques, achieving 100% accuracy on the KTH dataset.