Three-dimensional parameters tracking in multiview video sequences

Coherent immersion of real objects in virtual environment requires to know their spatio-temporal behaviour. In this paper we present a parameters tracking method for moving objects. This method uses multiple video streams of the tracked objects as input. In our approach we assume given a parametered model of the tracked subject, which includes geometrical description and parameters which are freedom degrees of the subject. The method solves the parameters tracking problem by finding the vector parameter which generates the closest synthetic views of the subject regards to the real images. This is done by introducing an adequation measure function between images using dense comparisons, allowing reliability and occlusions robustness. Parameters tracking is achieved using an adapted simplex-based algorithm, which gives fast and reliable results. We also present model auto-refinement. which allows to use an imprecise model of the tracked object, which is refined to match reality. Textures auto-extraction allows to extract real textures of objects, in order to improve model precision and realistic, or for analysis purpose. Experiments on real video sequences are presented. They show the efficiency and the robustness of the approach in various situations.