Unsupervised 3D human pose recognition from a single depth human silhouette using a geodesic map and kinematic body model

Abstract

Recently, human pose recognition (HPR) in 3D using only a single depth sensor without any optical markers has become an active research topic. Till now, most existing HPR approaches are based on supervised recognition of human body parts, requiring a classifier trained with a proper database. In this paper, we propose a novel unsupervised 3D HPR utilizing a geodesic distance map (GDM) of human depth silhouette and a 3D kinematic body model which requires no training and database. From each GDM, we derive a set of landmarks of human body joints and fit the joint landmarks of a kinematic body model to them to reconstruct its corresponding pose in 3D. Our numerical evaluation results of our proposed methodology indicate a range of errors from 0.01 to 33.45 mm in the Euclidean distance of the joints to their true location in 3D. Experimental results with real data demonstrate that the proposed technique could perform HPR in 3D with a reasonable accuracy and reliability.