Recognition, Tracking, and Reconstruction of Human Motion

Abstract

This paper describes a system aimed at automising the reconstruction of human motion. Human motion can be described as a sequence of 3D body postures. View based recognition of these postures forms the basis of human tracking algorithms [18]. These postures are defined by the underlying skeleton, an articulated structure of rigid links connected at rotational joints. The skeleton can be reconstructed if the rotational joints are tracked [11]. A set of posture specific key frames with pre defined joint locations are stored. Joint locations from these key frames can be mapped to actual frames once correspondence between the two shapes has been achieved. The rotational joints are in general not well defined in 2D images thus the iterative process of successively repeating point localisation and 3D reconstruction allows one to impose the geometric definition on the points. The power of the approach presented is demonstrated by the recognition, self calibration and 3D reconstruction of a tennis stroke seen from two cameras achieved without precalibrated cameras or manual intervention for initialisation and error recovery.