Tracking Multiple Articulated Objects Using Physics Engines: Improvement Using Multi Scale Decomposition and Quadtrees

This paper presents a new method to accelerate the automatic tracking of multiple articulated objects in video sequences. The tracking method exploits the physics modeling of the articulated objects. The physics approach has numerous advantages such as the capability to let multiple models interact when tracking simultaneously multiple objects. Solving the tracking problem is generally done with an iterative scheme involving the computation of numerous forces, which usually turns out to be time consuming. This paper tackles the computational cost bottleneck by introducing a multi scale decomposition of the force maps. At each step of the tracking optimization scheme the multiple force maps are efficiently combined using a quadtree structure to recover the resulting forces. We demonstrate through multiple experiments that our results are both exact and drastically faster than the standard computation scheme.