Multi-level Sparse Coding for Human Action Recognition

Sparse coding is a popular feature coding method in human action recognition, but the feature representation constructed under sparse coding framework cannot capture meaningful contextual information of local features. To address this problem, we propose a multi-level sparse coding method. Concretely, we defined several contexts for each local feature to capture the spatio-temporal contextual information in multiple structures and scales, and descript each context by max pooling the coding vectors in the context, then construct multiple vocabularies. The experimental results evaluated on KTH and YouTube datasets reveal that our method achieves state-of-the-art performance.