Model-based analysis of multi-video data

Abstract

We describe a method to analyze multiple synchronized video streams by making use of a parameterized geometry model of the recorded object. By formulating the task of fitting the model to the images in terms of optimizing silhouette match, we are able to capture the time-varying 3D pose of the object automatically and robustly. To evaluate the energy functional, we exploit the fast image synthesis capabilities of a conventional PC graphics card. The use of an a-priori object model enables us to enforce kinematic constraints as well as temporal coherence, and we obtain a high-quality surface description as output. Suitably modified, the presented technique is also applicable to medical and other image analysis tasks if a parameterized, generic geometry model of the object of interest is available.