Unsupervised human action retrieval using salient points in 3D mesh sequences

Abstract

The problem of human action retrieval based on the representation of the human body as a 3D mesh is addressed. The proposed 3D mesh sequence descriptor is based on a set of trajectories of salient points of the human body: its centroid and its five protrusion ends. The extracted descriptor of the corresponding trajectories incorporates a set of significant features of human motion, such as velocity, total displacement from the initial position and direction. As distance measure, a variation of the Dynamic Time Warping (DTW) algorithm, combined with a k − means based method for multiple distance matrix fusion, is applied. The proposed method is fully unsupervised. Experimental evaluation has been performed on two artificial datasets, one of which is being made publicly available by the authors. The experimentation on these datasets shows that the proposed scheme achieves retrieval performance beyond the state of the art.