Motion recognition using local auto-correlation of space–time gradients

In this paper, we propose a motion recognition scheme based on a novel method of motion feature extraction. The feature extraction method utilizes auto-correlations of space–time gradients of three-dimensional motion shape in a video sequence. The method effectively exploits the local relationships of the gradients corresponding to the space–time geometric characteristics of the motion. For recognizing motions, we apply the framework of bag-of-frame-features, which, in contrast to the standard bag-of-features framework, enables the motion characteristics to be captured sufficiently and the motions to be quickly recognized. In experiments on various datasets for motion recognition, the proposed method exhibits favorable performances as compared to the other methods, and faster computational time even than real time.

► We propose a novel motion feature extraction method, called STACOG.
► An efficient framework using STACOG features for recognizing actions is presented.
► The experimental results on various action datasets exhibit favorable performances.
► Faster computation time (even than real time) for action recognition is shown.