Tile-Based Mesh Refinement for Large 3D Terrain Visualization

• Main Authors: Huan-chih Chiu, 邱煥智
• Other Authors: Fuan Tsai
• Format: Others
• Language: zh-TW
• Published: 2008
• Online Access: http://ndltd.ncl.edu.tw/handle/sw27u6

碩士 === 國立中央大學 === 土木工程研究所 === 96 === Three-Dimensional (3D) terrain rendering is one of the most important components in the visualization of cyber city and other 3D GIS applications. When dealing with large-area terrain visualization, the vast amount of data may exceed the rendering capability of graphic hardware and cause poor performance of the system. Most importantly, in real-time visualization applications, the data resolution is much higher than screen, thus resulting in data redundancy and lowering efficiency. Furthermore, it may produce aliasing artifacts when rendering dense meshes. In order to reduce the number of polygons, tile-based approaches have become popular for large-area terrain visualization because the original DEM data can be pre-processed by tiles and only visible tiles need to be rendered in runtime. When rendering, the data will be loaded and rendered quickly without further effort for triangulation. In a previous study, a set of LODs was generated for each tile using a dynamic quadtree algorithm. When rendering the terrain, view frustum culling was used to decide visible tiles and computed view importance to assign suitable tile LOD. This study further improves the developed tile-based terrain visualization system. Firstly, the original thresholds for quadtree were determined by the height difference in each tile, but it is difficult to connect the LODs with view importance. A new thresholding scheme based on view-dependent image-space error metric is proposed to achieve more reasonable LOD generation. Secondly, conventional LOD systems often divide the data set into small tiles geographically. This may result in poor performance or abrupt LOD changes when dealing with large-area visualization projects, especially during the initial stage. To address this issue, a discontinuous LOD system is developed to create a finer set of LOD tiles and a coarser LOD set for quick representation of large areas. The switch between the two LOD sets is established according to viewer altitude, distance and resolution dependency etc. Thirdly, when visualizing a terrain by mesh tiles, there are usually discontinuities along tile edges, causing so-called T-junctions among different tiles. A mesh-merging algorithm is also proposed to refine the determined LOD meshes in order to eliminate T-junctions. Augmented with these improvements, the developed system will produce seamless landscape scenes consisting of multiple tiles of different LOD layers more efficiently. The developed tile-based terrain rendering system allows users to obtain near real-time visualization of large terrain data sets. The proposed new thresholding scheme based on Ground Sample Distance enables more reasonable quadtree-based LOD generation for data pre-processing. It also provides better relationship between the LOD generation and view-importance for determining appropriate LOD levels of visible tiles. The developed T-junction removal algorithm can eliminate discontinuities between adjacent tile meshes effectively and have little impact to the overall rendering performance. Using discontinuous LOD improves the performance significantly, especially during the initial stage of visualization. Test examples with two large DEM datasets conducted in this study demonstrate that the developed system can produce seamless rendered scenes with high performance in near real-time visualization applications.