Representing human motion with FADE and U-FADE: an efficient frequency-domain approach

In this work, we present FADE, a frequency-based descriptor to encode human motion. FADE is simple, and provides high compression rate and low computational complexity. In order to reduce space and time complexity, we exploit the biomechanical property that human motion is bounded in frequency. FADE and U-FADE can be used in combination with both supervised and unsupervised learning approaches in order to classify and cluster human actions, respectively. We present also a branch of FADE, called Uncompressed FADE (U-FADE). U-FADE performs well in combination with some unsupervised algorithms such as spectral clustering, paying the price of a reduced compression rate. Also, U-FADE performs in general better than FADE well with small datasets. We tested our descriptors with well-known, public motion databases, such as HDM05, Berkeley MHAD, and MSR. Moreover, we compared FADE and U-FADE with diverse state of the art approaches.