An X-T slice based method for action recognition

This paper proposes a novel method for human action recognition. Different from many action recognition methods which consider an action sequence along the time axis, the proposed method views an action sequence along the space axis. This brings two advantages: the human body structures in all frames are encoded in the feature; the time information is completely used. The process of feature extraction is as follows: first an action sequence is cut into slices parallel to the X-T plane. Every slice, we call X-T slice, is transformed to a mean histogram and a variance histogram along the T axis. Then all mean histograms and all variance histograms are concatenated separately to two vectors, and finally encoded with Mel Frequency Cepstrum Coefficient (MFCC). MFCC, a feature commonly used in speech recognition, can effectively capture changes of 1-D signals over time. The encoded values are sent to classifier for action recognition. Our system achieves very efficient result: it needs only 0.02 second to deal with a frame on average with Matlab.