Tracking hybrid 2D-3D human models from multiple views

Abstract

A novel framework is proposed under which robust matching and tracking of a 3D skeleton model of a human body from multiple views can be performed We propose a method for measuring the ambiguity of 2D measurements provided by each view. The ambiguity measurement is then used for selecting the best view for the most accurate match and tracking. A hybrid 2D-3D representation is chosen for modelling human body poses. The hybrid model is learnt using hierarchical principal component analysis. The CONDENSATION algorithm is used to robustly track and match 3D skeleton models in individual views.