A new Approach to Speed up in Action Recognition Based on Key-frame Extraction

Human action recognition is the process of labelingvideos contain human motion with action classes. The run timecomplexity is one of the most important challenges in actionrecognition. In this paper, we address this problem using videoabstraction techniques including key-frame extraction and videoskimming. At first we extract key-frames and then skim the videoclip by concatenating excerpts around the selected key-frames.This shorter sequence is used as input for classifier. Ourproposed approach not only reduces the space complexity butalso reduces the run time in both train and test steps. Theexperimental results provided on KTH action datasets show thatthe proposed method achieves good performance without losingconsiderable classification accuracy.

Keywords:

action recognition , key-frame , unsupervised feature extraction , KTH dataset