A spatio-temporal interest point detector based on vorticity for action recognition

How to separate foreground from camera and background motions is a difficult problem for human action recognition in unconstrained environments. Although the existing interest point based methods have shown attractive results, they always come with high computational complexity and lose their power in cluttered field with camera motion. In this paper, a new spatio-temporal interest point detector based on flow vorticity is proposed, which can not only suppress most of effects of camera motion but also provide prominent spatio-temporal interest points around key positions of the moving foreground. Experiments on KTH and UCF sports datasets demonstrate that the proposed detector combined with HOG/HOF descriptors outperforms four popular point detectors. Moreover, when rich descriptors, such as trajectory, HOG/HOF and MBH (Motion Boundary Histogram), are combined with the proposed detector, our approach offers better results than other state-of-the-art methods, and also achieves comparable performance with less than half the computation time compared with dense sampling. Our approach is simple and effective, and manifests a good tradeoff between recognition accuracy and computational complexity.