Human action recognition using sparse representation

Sparse representation has been applied recently to many signal processing and computer vision and demonstrated successful results. Inspired by them, we propose an action recognition approach based on sparse representation to avoid the sensitivity of parameter selection in nearest-neighbor classification method and improve the discriminative capability. Firstly, each frame in the test sequence is treated as a sparse linear combination of all frames in the training sequences, and its sparsest representation is computed by L1-minimization. Then each frame is classified by minimizing the residual. Finally, we classify the testing sequence based on the majority of these frames' classes. Experiments are conducted on two publicly availabe datasets: Weizmann dataset and IXMAS multiview dataset. The results demonstrate that our approach achieves better performance than nearest-neighbor, and outperforms most recently proposed methods.