We present a method to perform occlusion culling on moving objects in a 3D scene. This method computes potentially visible sets of polygons (PVS) from cells that are based on the bounding box of the object and have the particularity to be overlapping. The visibility is precomputed from several renderings of the object. The implementation is easy, with a data structure that does not require additional memory compared to the original data used to describe the object. It provides interesting acceleration and smooth animation even when the list of polygons sent to the display device has to be updated. Since it is an aggressive method, we discuss the errors that may occur and the ways to fix them.