Abstract: We propose a novel hardware-accelerated voxelization algorithm for polygonal models. Compared with previous approaches, our algorithm has a major advantage that it guarantees the conservative correctness in voxelization: every voxel intersecting the input model is correctly recognized. This property is crucial for applications like collision detection, occlusion culling and visibility processing. We also present an efficient and robust implementation of the algorithm in the GPU. Experiments show that our algorithm has a lower memory consumption than previous approaches and is more efficient when the volume resolution is high. In addition, our algorithm requires no preprocessing and is suitable for voxelizing deformable models.