Human Action Recognition Using Key Points Displacement

Recognizing human actions is currently one of the most active research topics. Efros et al. first proposed using optical flow and normalized correlation to recognize distant actions. One weakness of the method is that optical flow is too noisy and cannot reveal the true motions; the other popular method is the space-time-interest-points proposed by Laptev et al., who extended the Harris corner detector to temporal domain. Inspired by the two methods, we proposed a new algorithm based on displacement of Lowe's scale-invariant key points to detect motions. The vectors of matched key points are calculated as weighted orientation histograms and then classified by SVM. Experimental results demonstrate that the proposed motion descriptor is effective on recognizing both general and sport actions.