High-Speed Action Recognition and Localization in Compressed Domain Videos

We present a compressed domain scheme that is able to recognize and localize actions at high speeds. The recognition problem is posed as performing an action video query on a test video sequence. Our method is based on computing motion similarity using compressed domain features which can be extracted with low complexity. We introduce a novel motion correlation measure that takes into account differences in motion directions and magnitudes. Our method is appearance-invariant, requires no prior segmentation, alignment or stabilization, and is able to localize actions in both space and time. We evaluated our method on a benchmark action video database consisting of six actions performed by 25 people under three different scenarios. Our proposed method achieved a classification accuracy of 90%, comparing favorably with existing methods in action classification accuracy, and is able to localize a template video of 80 x 64 pixels with 23 frames in a test video of 368 x 184 pixels with 835 frames in just 11 s, easily outperforming other methods in localization speed. We also perform a systematic investigation of the effects of various encoding options on our proposed approach. In particular, we present results on the compression-classification tradeoff, which would provide valuable insight into jointly designing a system that performs video encoding at the camera front-end and action classification at the processing back-end.