Model-based video tracking for gestural interaction

摘要

Among many techniques to interact with 3D environments, gesture-based input appears promising. However, due to insufficient computing hardware capabilities, such interfaces have to be built either upon standard tracking devices or using limited image-based video tracking algorithms. As today computing power tends to be more and more powerful, more complex video analysis such as real-time model-based tracking is at hand. Considering the use of a model-based approach to allow unencumbered input gives us the advantage of extracting a low-level hand description useful to build natural interfaces. The algorithm we developed relies on a 3D polygonal hand model. Its pose parametrization is iteratively refined so that its 2D projection matches more closely the input 2D image. Relying on the graphics hardware to handle fast 2D projection is critical, while adding more cameras is useful to cope with the occlusion issue.