Available online at www.sciencedirect.com

ScienceDirect

Procedía Earth and Planetary Science 1 (2009) 1147-1151

Procedia Earth and Planetary Science

www.elsevier.com/locate/procedia

The 6th International Conference on Mining Science & Technology

View-dependent fast real-time generating algorithm for

large-scale terrain

Jin Hai-liang^*, Lu Xiao-pinga, Liu Hui-jiea

aKey Laboratory of Mine Spatial Information Technologies of State Bureau of Surveying and Mapping, Henan Polytechnic University, Jiaozuo

454000, China

bKey Laboratory of Advanced Design and Intelligent Computing, Dalian University, Dalian 116622, China

Abstract

In order to eliminate the popping effect of switching among levels of detail and to increase the frame rates with high image quality during the fly-through of large-scale terrain, this paper put forward a new bottom-up modeling strategy, which constructed simplified terrain triangle mesh globally and updated mesh nodes dynamically. Employed hybrid culling technique based on blocks and triangle faces and simplified computing method for screen-space errors to select appropriate terrain nodes rapidly. Then updated the Delaunay terrain mesh by adding nodes, deleting nodes and modifying locally. At the same time achieved self-adaptive control for screen-space error tolerance during the terrain fly-through. Results of simulation experiments demonstrate that the algorithm eliminates popping effect effectively, and has a higher frame rate compared with other algorithms. So it is particularly suitable for close-distance fly-through simulation of large-scale terrain.

Keywords: view-dependent; real-time visualization; large-scale terrain

1. Introduction

In recent years, large-scale terrain fast real-time rendering is a hot research topic in relative field. One of the core problems of terrain visualization research is to resolve the conflict between limited rendering ability of computer graphics hardware and complex terrain surface model constructed by massive terrain data. It is for this reason that someone has put forward the level of detail technology (LOD) to simplify the surface model and improve rendering efficiency. Current LOD technology can be divided into two types: view-independent and view-dependent LOD algorithm. View-independent LOD algorithms are simple, and easy to implement, but such algorithms need to store a lot of model independent each other, and put forward higher requirements for the computer's hardware. In addition, there would cause popping phenomenon because that the transition between different resolution models is not quite natural. Most of the view-dependent LOD algorithms are using binary tree or quadtree to subdivide and simplify terrain model based on blocks [1-4]. Hoppe applied view-dependent progressive mesh model (VDPM) to

* Corresponding author. Tel.: +86-13261535175. E-mail address: jin_hailiang@126.com.

1878-5220/09/$- See front matter © 2009 Published by Elsevier B.V. doi:10.1016/j.proeps.2009.09.176

real-time rendering of terrain [5], and adopt geomorphing to eliminate popping between different levels. Such algorithm used different resolution levels between the terrain blocks, easily lead to cracks, need for additional measures to eliminate the cracks. At the same time, frustum culling based on grid-vertex and triangle requires substantial computation with more intensive sampling points, which reduce the efficiency of the algorithm to some extent. On the basis of existing research results, this paper put forward a new real-time rendering technology of large-scale terrain with the idea of LOD, using block-based frustum culling, structuring mesh globally at one-time, and using real-time update strategy for mesh nodes, effectively improving the structure mesh speed of the terrain model. We avoid the problems of eliminating cracks with the consistency of the resolution at structuring mesh. Finally, the authors have applied simulation experiment using a terrain data as an example.

2. Subdivision basis of TIN model of large-scale terrain

The algorithm of this paper selects triangulated irregular network (TIN) as the data model of terrain simulation. In this paper, we used Bowyer and Watson's incremental algorithm [6] (empty algorithm) to generate Delaunay triangulation mesh. The basic principle is as follows: First of all, to generate a big triangle TO which contains all the points of the given point set; for each vertex P of the given point set, to find all those triangle which circumcircle contain points P, and note down the region border composed by these triangles (Delaunay empty hole), and delete these triangles; then connect point P and every edge of the empty border to form new triangles, and joined these triangles into the triangular mesh; finally, to find those triangles which vertexes include vertexes of the triangle T0, and deleted those triangles; algorithm end.

The main computing of above-mentioned subdivision process includes two aspects: a) to find all triangles which circumcircle contain points P; b) connect point P and every edge of the Delaunay empty border to form new triangles. In order to speed up the search speed of algorithm, this paper embeds compass algorithm [7] into the above-mentioned algorithm; the efficiency of algorithm is close to O (nlgn) in this way, n is of the number of points of point set.

3. Algorithm design and implementation

This paper presents an algorithm design concept which is structuring mesh globally at one-time and using realtime update strategy for mesh nodes. The algorithm maintains only one limited points set that meet the requirements of the screen-space error and are view-dependent in the whole process of algorithm, and generates terrain triangulation subdivision mesh using this points set by Delaunay incremental algorithm at one-time; then updates the limited points set according to the change of view parameters, adopts adding point and deleting point operation for generated triangular mesh to achieve view-dependent dynamic change of terrain details. The whole algorithm process can be divided into two parts, namely, constructing the initial terrain mesh model and real-time dynamic update of terrain mesh model, in which a) - c) is rendering process of the first frame terrain mesh model.

a) Blocking pretreatment of terrain DEM data. The algorithm divides DEM data into m*n sub-block using average blocking method, establishes spatial index data structure of sub-block, computes and saves radius of surrounded ball and elevation standard deviation of each sub-block.

b) Frustum culling. Using a simplified cutting algorithm, calculates sub-block set a = U block intersect with frustum by radius of the surrounded ball.

c) According to the initial view parameters and pre-set screen-space error limit, selects node meeting the requirements from the set a, generates view-dependent limited point set P, and generates simplified terrain model of point set P by Delaunay incremental subdivision algorithm.

d) Real-time update of mesh nodes. According to the change of view parameters and screen-space error limit, real-time calculates mesh nodes entering and leaving frustum, carry out adding point and deleting point operation for generated terrain Delaunay triangular mesh by c), and dynamic updates mesh simplified model.

e) Sent real-time generating final terrain simplified model of each frame to 3D graphics rendering engine, singleframe algorithm loop end.

It is easily to find that this paper uses a bottom-up simplified strategy different from the general quadtree terrain simplification algorithm, which generates simplified model gradually by the fine model. The algorithm structures terrain simplified mesh model globally at one-time only in c), from d) begun to enter dynamic loop update process

of terrain model. In this process, there is no need to re-structure mesh, only to have carried out adding point and deleting point operation for generated mesh model, and realize real-time updates of simplified model. Due to the simplified model of each frame is a partial update for last frame, rather than conversion between different resolution layers in algorithm, effectively overcome the popping phenomenon generated in terrain fly-through; at the same time, due to the consistency of resolution at constructing mesh, avoid fundamentally the appearance of cracks, effectively improve the efficiency of the algorithm.

3.1. DEM data blocking

This paper used data block strategy in order to achieve mainly fast block-based frustum culling, at the same time, to realize easily the spatial location of the model surface nodes. Using the average block method, DEM data is divided into m x n sub-blocks, the size of each sub-block satisfy (2l +1) x (2l +1), if not satisfied, the sub-block can be filled out with gray data points (elevation of gray data points is an impossible constant value, not to participate in operation of constructing mesh). Radius of surrounded ball of each sub-block and elevation standard deviation are calculated and saved. Radius of surrounded ball of each sub-block is used for frustum culling and elevation standard deviation is used as a control factor for screen-space error limit.

3.2. Selecting mesh nodes for model simplification

Main consideration of the algorithm is how to select suitable terrain vertex, making simplified mesh model approaching greatest possible to the original terrain model. This paper reference to terrain vertex selection method in reference [2], and modify the method to meet the needs of the algorithm. The thought of selecting mesh nodes is: First of all, define the height difference of each vertex according to the vertical height difference of surrounding topography of terrain vertex, and then projected the height difference of vertex onto the screen coordinate system. If the projection of the height difference is greater than given allowable value of error, this vertex should be selected. We defined the height difference of vertex firstly, and then explain how to calculate projection of the height difference of terrain vertex.

This paper uses variation of slope between co-triangle pair to measure height difference of terrain vertex. As shown in Fig. 1., terrain vertex B is the shared vertex of co-triangle pair AABE and ABCE of mesh model. When the co-triangle pair combined into a triangle A ACE, terrain vertex B is selected or do not selected according to the value of slope of the plane which co-triangle pair lies on transfers to the plane which triangle AACE lies on respectively. In practical application, when terrain vertex has the same height difference, but the distance to viewpoint or the angle between projection vector (the vector that terrain vertex connected with viewpoint called projection vector) of terrain vertex and line-of-sight direction is not the same, the changes in visual caused by triangle merged simplification is not the same. So, this paper selects terrain vertex which satisfies requirements according to the projection of height difference, and thus structure the final simplified model of terrain meshes. Fig. 2. shows the projection calculation principle of height difference.

In the Fig. 2., the midpoint of height difference line in the world coordinates system (WCS) is v, its coordinates are (vx, vy, vz); 5 is the height difference of terrain vertex; e is viewpoint; d is the distance from viewpoint to view plane; parameter X is the conversion constant between the length units of WCS and the length units (pixels) of SCS (screen coordinate system); then the screen projection of height difference can be calculated as follows:

Fig. 1. Sketch map of calculating height difference

Fig. 2. The projection of height difference

The decision condition of terrain vertex is 5screen(e, v)>t. t is the pre-set projection limit of height difference, using pixel units of screen coordinate system. The projection limit of height difference indicates that the error of the imaging effect between established triangular meshes using selected terrain vertex according to this criteria and the original terrain model will not be greater than t pixels on the screen.

3.3. View-dependent mesh model dynamic update

There are two crucial factors for a high-performance real-time fly-through simulation algorithm of large-scale terrain: elimination of image popping phenomenon and higher frame rate of graphics rendering. For the first factor, this paper uses the strategy of structuring mesh globally at one-time and real-time update mesh nodes, and there not exist transition problem between different details models, effectively avoid image popping phenomenon when terrain fly-through. For the second factor, this paper adopts adding point and deleting point operation based on Delaunay triangulation mesh to achieve dynamic updating of terrain model when terrain fly-through. The algorithm is simple and low computational complexity, and ensures a high frame rate for graphics rendering. Terrain blocks enter or leave frustum constantly in the process of viewpoint moving. For the terrain block entered into frustum, selecting the suitable terrain node according to the projection limit height difference, and adds to the current terrain triangular meshes. For the terrain block left frustum, deletes directly the corresponding nodes in the current terrain triangular meshes. For the terrain block still inside frustum, it is need to re-calculate the projection of height difference because that the change of viewpoint cause the change of the projection of height difference, and then update meshes (including adding point and deleting point operation) according to the current value of t. The process of real-time dynamic updates of mesh model is shown in Fig. 3.

Jf _, T . • J deUtiitg ..... k/ node J

nillirl

\ : , , >ii(.n>h»

loute of vicwpmju kiovulj •

Fig. 3. Dynamic updates of mesh model

In the course of terrain fly-through, the authors found that it can improve the visual effects of graphics fly-through by lowering appropriately the value of t in the region where terrain is more complex and changes rapidly and raising appropriately the value of t in the region where terrain changes is comparatively small(flat). At the same time, in order to maintain a high frame rate of graphics rendering, it is needed to adjust the value of t according to the speed of viewpoint moving, raising appropriately the value of t when the speed is fast (using rough terrain model), on the other hand, e lowering appropriately the value of t when the speed is slow (using fine terrain model). Therefore, this paper has established the experience function mapping relations between the projection limit t of height difference and viewpoint moving speed v, and the complexity of the terrain (denoted by standard deviation c of elevation value of block), as following:

T = ct + 1/(1+ CT + c2 /1 v |)

In the formula, ci is a constant indicating the number of pixels, and control the value of t at the macro level; c2 is an experience constant relevant to viewpoint moving speed; the latter part of formula achieve self-adaptive micro-control of the projection limit t of height difference.

4. Simulation experiment and analysis

In order to verify the validity of the rendering algorithm for large-scale terrain that this paper puts forward, the author has carried out simulation experiment using multi-group terrain DEM data on a 2.4 GHz Pentium IV computer with 512 MB memory and GeForce 4Ti4600 graphics card with 128 MB video memory; 3D graphics standard is OpenGL. Terrain DEM data with size of 2 049 x 2 049 is divided average into 32*32 sub-blocks, each

block has 65x65 sampling points. Three-dimensional effect of terrain fly-through shows in Fig. 4.

Fig. 4. Three-dimensional effect of terrain fly-through

In the above simulation experiment, viewpoint moving speed is controlled in the range 70 ~ 100 m / s. we control the frame rate and the number of triangles of real-time rendering by changing the projection limit t of height difference. The average frame rate is 36 fps on condition that the average value of t is 2.5 and the average number of triangles is 15000.

5. Conclusions

Eliminating the phenomenon of image popping and maintaining a higher frame rate of graphics rendering is always two major problems need to solve for visualization simulation algorithm of large-scale terrain. This paper put forward a new strategy of structuring mesh globally at one-time and meshes nodes real-time updating, and achieved good simulation results in this the two sides, especially when close-distance fly-through simulation, the effect is more desirable. Further work for the algorithm of this paper mainly include resolving the optimal problem of calculating the projection limit of height difference, as well as the compression problems of ultra-large-scale terrain data import directly into memory.

Acknowledgements

This work is supported by National Basic Research Program of China N0.2009CB226107, Key Laboratory of Mine Spatial Information Technologies of State Bureau of Surveying and Mapping (Henan Polytechnic University) N0.KLM200912, Doctorial Fund of Henan Polytechnic University NO. B2009-23, Key Laboratory of advanced Design and Intelligent Computing (Dalian University) N0.2007-1

References

[1] M. Duchaineau, M. Wolinsky and D.E. Sigeti, Roaming terrain: real-time optimally adapting meshes, Proc of IEEE Visualization, 1997.

[2] P. Lindstrom and D. Koller, Ribarskyw, Real-time, continuous level of detail rendering of height fields, Proc of SIGGRAPH'96. Los Angeles, California, 1996.

[3] Y. Zhao, J. Shi and J. Zhou, A Fast Algorithm for Large Scale Terrain Walkthrough, Journal of Computer-Aided Design & Computer Grahhics, 14 (2002) 624-628.

[4] Y. Lu, H. Jiang and B. Tang, View-Dependent Simplification Algorithm for Complex Models. Mini-Micro Systems, 26 (2005) 156-160.

[5] Hoppeh, Smooth view-dependent level-of-detail control and its application to terrain rendering, Proc of IEEE Visualization, 1998.

[6] Q. Yang, Conforming Delaunay Triangulation technology. Beijing:Publishing House of Electronics Industry, 2005.

[7] Z. Li, The study on general visualization techniques for three dimensional scattered data field in broad sense domain. Beijing: Beijing University of Aeronautics & Astronautics, 1997.

[8] Y. Wu, Y. Du, X. Wang, Lod-based Algorithm of Tin Model for Large Terrain Simulation. Journal of System Simulation. 17 (2005) 665669.