We present an output-sensitive occlusion culling algorithm for densely occluded dynamic scenes where both the viewpoint and objects move arbitrarily. Our method works on a regular grid that represents a volumetric discretization of the space and uses the opaque regions of the scene as virtual occluders. We introduce new techniques of efficient traversal of voxels, object discretization and occlusion computation that strengthen the benefits of regular grids in dynamic scenes. The method also exploits temporal coherence and realizes occluder fusion in object-space. For each frame, the algorithm computes a conservative set of visible objects that greatly accelerates the visualization of complex dynamic scenes. We discuss the results of a 2D and 3D case implementation.