Optical flow-motion history image (OF-MHI) for action recognition

摘要: The motion history image (MHI) is a global spatiotemporal representation for video sequences. It computationally very simple and efficient. has been widely used many real-time action recognition tasks. However, the conventional MHI assigns fixed strength to each detected foreground point then updates it with small constant background point. Local body parts different movement speeds durations will have same intensity in MHI. Similar actions may generate indistinguishable patterns. In this paper, we propose new that incorporates both optical flow revised of pixel adaptively accumulated by length at location. exponentially updated over time. can better describe local movements temporal template. duration implicitly given update rate description various scene. For classification, set training samples are first collected form basis templates. An sequence constructed as linear combination coefficients give feature vector. Euclidean distance finally evaluate similarity between vectors. Experimental results on KTH Weizmann datasets shown proposed scheme yields 100 % rates test fast processing 47 fps $$200\times 150$$ images.