Contextual Fisher Kernels for Human Action Recognition

In the literature of human action recognition, despite promising results have been obtained by the traditional bag-of-words model, the relationship among spatio-temporal points has rarely been considered. Furthermore, serious quantization error also exists in this kind of strategy. In this paper, we propose a novel coding strategy named contextual Fisher kernels to overcome these limitations. We add a Gaussian function to represent contextual information into the generative model. In this way, our method explicitly considers the spatio-temporal contextual relationships between interest points and alleviates quantization error. Our method is verified on two challenging datasets (KTH and UCF sports), and the experimental results demonstrate that our method achieves better results than the state-of-the-art methods in human action recognition.