Abstract: We present a collision detection (CD) method for complex and large-scale fracturing models that have geometric and topological changes. We first propose a novel dual-cone culling method to improve the performance of CD, especially self-collision detection among fracturing models. Our dual-cone culling method has a small computational overhead and a conservative algorithm. Combined with bounding volume hierarchies (BVHs), our dual-cone culling method becomes approximate. However, we found that our method does not miss any collisions in the tested benchmarks. We also propose a novel, selective restructuring method that improves the overall performance of CD and reduces performance degradations at fracturing events. Our restructuring method is based on a culling efficiency metric that measures the expected number of overlap tests of a BVH. To further reduce the performance degradations at fracturing events, we also propose a novel, fast BVH construction method that builds multiple levels of the hierarchy in one iteration using a grid and hashing. We test our method with four different large-scale deforming benchmarks. Compared to the state-of-the-art methods, our method shows a more stable performance for CD by improving the performance by a factor of up to two orders of magnitude at frames when deforming models change their mesh topologies.