The dynamics of linear combinations: tracking 3D skeletons of human subjects

We propose a general framework for addressing three fundamental issues using linear combinations: (1) the properties of the examples to linearly combine, (2) the constraints, and (3) the method for estimating the linear combinations coefficients for reconstructing an object based on noisy and incomplete visual observations. To this end, we synthesise the necessary examples from known data using principal component analysis. Crucially, the dynamics of the object is dealt with by learning spatio-temporal constraints on the coefficients of the linear combinations. The CONDENSATION framework was adopted to estimate the coefficients for legitimate and plausible linear combinations. Finally, we apply the linear combinations framework to track 3D skeletons of human subjects using a hybrid representation. (C) 2002 Published by Elsevier Science B.V.