A novel, view-dependent rendering method integrated with occlusion culling for many articulated models used in large-scale crowd scenes is presented. A cluster hierarchy for an animated, articulated model, is used as a dual representation for view-dependent rendering and occlusion culling. For high-performance view-dependent rendering and occlusion culling, each cluster of the cluster hierarchy is constructed to contain a spatially coherent portion of the mesh that also has similar simplification errors, by using our error-aware clustering method. We perform the well-known quadrics-based simplification to efficiently compute our representation, while achieving a high quality simplification. At runtime, we choose a LOD cut from the cluster hierarchy given a user specified error bound in the screen space and render all the visible clusters in the LOD cut. We also perform occlusion culling by utilizing the temporal coherence among consecutive frames. We implement our method in GPU and achieve interactive performance (e.g., 50 frames per second) for large-scale crowd scenes that consist up to one thousand articulated models and 83 M triangles, without noticeable visual artifacts. Compared to a base rendering system that performs only view-frustum culling, our method achieves up to 6 times improvements in our benchmark crowd scenes.