Human action recognition in the fractional Fourier domain

Most studies about silhouettes based human action recognition focus on the time domain representation. However, the contour of human body usually shows as a time-varying signal, for which neither the time domain based methods nor the Fourier transform can catch enough information to achieve sufficient classification performance. A fractional Fourier shape descriptor is proposed for silhouette based human pose representation and action recognition. The fractional Fourier shape representation of human silhouette is more robust and discriminative than that in the time or frequency domain. A criteria called diffusion score is proposed to determine the best fractional order. After the fractional shape features are built, we propose a two-stage random forest based framework to classify human poses in the action sequence and vote the action label. Experimental results on benchmark dataset show that our method is effective.