An efficient collision detection method based on separating bounding volume (SBV) is proposed. The positions and shapes of SBVs are determined by the optimal separating support hyper planes of two objects. SBVs not only can efficiently detect the separation of models, but have a high culling ratio when models are intersecting. In order to compute SBVs efficiently, an approximate method using SVM is also put forward and tested. At last in penetration region, a method combined with GPU and SBVs is designed to handle the proximity queries. Experimental results illustrate that SBVs based collision detection algorithm is applicable to exact collision detection for 3D models even without topologies and achieves more efficient and balanced performances in separating, colliding and especially puncturing cases.