The observation of the direct vehicle surroundings is a critical task to both (semi-)autonomous vehicles and human drivers. A surround view from a virtual bird's view camera perspective helps to enhance operational safety in both cases. Yet, state-of-the-art methods for the computation of these views rely on planar ground plane assumptions which lead to systematic errors and highly distorted views especially on uneven ground or in off-road applications. We address this issue and propose an approach for the computation of dense perspectively correct surround views using stereo vision, a closed-surface heightmap and partial homographies. Using temporal integration of the stereo images, a high image quality is achieved.