Hierarchy Accelerated Stochastic Collision Detection

In this paper we present a new framework for collision and self-collision detection for highly deformable objects such as cloth. It permits to efficiently trade off accuracy for speed by combining two different collision detection approaches. We use a newly developed stochastic method, where close features of the objects are found by tracking randomly selected pairs of geometric primitives, and a hierarchy of discrete oriented polytopes (DOPs). This bounding volume hierarchy (BVH) is used to narrow the regions where random pairs are generated, therefore fewer random samples are necessary. Additionally the cost in each time step for the BVH can be greatly reduced compared to pure BVH-approaches by using a lazy hierarchy update. For the example of a cloth simulation framework it is experimentally shown that it is not necessary to respond to all collisions to maintain a stable simulation. Hence, the tuning of the computation time devoted to collision detection is possible and yields faster simulations.