Real-world large-scale terrain model reconstruction and real-time rendering

Abstract

We present an efficient framework for reconstructing real-world terrain models and achieving high-fidelity real-time rendering. In the terrain model reconstruction phase, we utilize Digital Elevation Model (DEM) or LiDAR point clouds as input data. For LiDAR point clouds, we introduce a normal estimation algorithm that incorporates LiDAR scan information. This enables us to generate an oriented point cloud, which serves as the basis for 3D reconstruction. The resulting mesh models can preserve landform features such as cliffs and overhangs, leading to highly realistic terrain models. To facilitate efficient rendering, we construct a Hierarchical Level of Detail (HLOD) representation based on the generated terrain models. The HLOD structure establishes a hierarchical relationship across successive LODs. During the rendering stage, we dynamically traverse the hierarchy and select the appropriate resolution tiles based on the viewpoint's position. This approach ensures smooth LOD transitions and eliminates issues such as LOD popping and mesh cracks through the use of vertex interpolation. We have implemented and tested our framework on a modest desktop machine without a discrete GPU. The HLOD construction implementation achieves a processing speed of 1 million triangles per second per core. During real-time rendering, our approach maintains an interactive frame rate without introducing artifacts. Our pipeline effectively leverages the capabilities of an average graphics card, ensuring efficient utilization of available resources.