Research On The Key Technologies For Regular-grid Real-time Rendering

With the development of modern mapping and remote sensing technology, the size ofelevation and image data is growing at an unprecedented rate. From digital city to digitalChina, and then to digital earth, terrain data are growing exponentially. At the same time,the requirement of realism and rendering speed of the geometric model in the virtual sceneis also growing. The regular grid data, focusing on elevation images and surfacerefinement templates, represent common characteristics, such as the large amount of data,the spatial layout of regular, and the need for real-time rendering.The real-time rendering of regular grid faces some technical difficulties, such as thelack of scientific and effective space error threshold setting for level of details(LOD), thelack of scientific quantization precision to guide the geometric data compression, the lowperformance of view visibility and LOD selection computing on Tiled Quadtree, the lackof efficient and flexible crack treatment for LOD, the low adaptivity of surface refinement,the difficulty to be extended for the mobile platform and so on. These restrict the scalableof data size, render quality and efficiency.Considering the lack of a reasonable space error threshold for LOD, the world-spaceerror bound was set for each level of Tiled Quadtree by referencing screen space Firstly.Secondly, we gave the optimization of vertex simplification and quantization bits,considering all kinds of errors introduced by terrain processing and rendering, andcombining the current function computing the data size of tile compressed mesh.Considering the lack of Tile Load Map, we proposed Multilevel Tile Load Mapalgorithm. We analyzed efficient architecture features of massive terrain LODvisualization, and found that CPU could hardly select tiles from mass terrain effectivelywithin the real-time limit. This restricts the expansion of terrain’s size. Multilevel TileLoad Map (MTLM) algorithm was proposed for tile selection to extend TLM. It uses2dtexture for saving Tiled Quadtree (TQT), and executes tile view-frustum culling andlevel-of-detail selection by GPU acceleration. The scalability constraint was avoided bydealing tile selection with high-performance image processing in TLM. In addition, optimization was presented at tile data organization and resource management.In the light of the current regular grid crack treatment efficiency, we proposedparameterization method for multi-scale regular-grid model repairing cracks. Functioncurve fitting was utilized to change the mesh layout and eliminate T-junctions caused byvarious LODs transition. Moreover, the parameterization cracks-repairing methodsupports multi-span LOD using curve function family excellently.In analysing the low adaptivity of semi-uniform adaptive patch tessellation, weproposed GPU-based crack-avoiding adaptive even tessellation. Based on theparameterization method for multi-scale regular-grid model repairing cracks, in order toimprove semi-uniform adaptive patch tessellation, even tessellation was proposed usingvertex fitting function to increase tessellation level styles. Vertex was shifted bycontrolling function on vertex shader to eliminate the boundary cracks. In addition,parameter region was divided to avoid floating point precision issue caused by directioninconsistency of function coordinate axes. Then, parametric mesh layouts were optimized,and both vertex and index could be reused between different tessellation levels.Lastly, we gave an adaptive refinement pipeline for mobile terminals.Based onadaptive even tessellation, we proposed an adaptive mesh refinement pipeline based onPN triangles. The geometric model implemented interactive rendering on the mobileplatform.The experiments show that our researches effectively improve the efficiency ofregular-grid real-time rendering, and the above research results have been applied torelated863projects.