scalable large, multi-resolution terrain real-time...

Outline DARTS lab Introduction SimScape Large terrain modeling CLOD visualization Conclusion

Scalable Large, Multi-Resolution TerrainReal-Time Modeling and Visualization for Surface

System Simulations

Steven Myint Abhinandan Jain Jonathan CameronChristopher Lim

Jet Propulsion Laboratory, California Institute of Technology

IROS 2011

http://dartslab.jpl.nasa.govc©2011 California Institute of Technology. Government sponsorship acknowledged.

S. Myint et al. JPL

Scalable Large, Multi-Resolution Terrain Real-Time Modeling and Visualization for Surface System Simulations


1 DARTS lab

2 Introduction

3 SimScape

4 Large terrain modeling

5 CLOD visualization

6 Conclusion

S. Myint et al. JPL



DARTS lab

Dynamics And Real-Time Simulation(DARTS)

EDL simulations (DSENDS)

Rover simulations (ROAMS)

Airship simulations

Robotic arm simulations

S. Myint et al. JPL



DARTS lab examples (near earth object)

S. Myint et al. JPL



DARTS lab examples (ejecta)

S. Myint et al. JPL



DARTS lab examples (tethering near small body)

S. Myint et al. JPL



DARTS lab examples (surface operations)

S. Myint et al. JPL



DARTS lab examples (power analysis)

S. Myint et al. JPL



DARTS lab examples (spacecraft)

S. Myint et al. JPL



Large terrain modeling and visualization

S. Myint et al. JPL



Large terrain use cases

Our real-time simulations often requirelarge terrain modeling/visualizationsupport.

EDL simulations

Rover simulations

S. Myint et al. JPL



Large terrain use cases (continued)

Centimeter-resolution terraindata

Billions of vertices

Gigabytes of data

Too much data to load all atonce (due to time andmemory constraints)

Too much data to render inreal time (30 fps)

S. Myint et al. JPL



SimScape terrain framework

Models DEMs, planets, and arbitrary meshes

C++ and Python APIs

Store data in HDF5 (Hierarchical Data Format) for fastrandom access

Import data from various data formats (PDS, ISIS, GeoTiff,etc.)

S. Myint et al. JPL



Digital Elevation Maps

Regular rectangular height data

Fast access without mucharithmetic

Useful for modeling relatively smallareas

Used extensively in roversimulations

S. Myint et al. JPL



Planets

Grid of height data in sphericalcoordinates

Useful for larger areas when wemust consider planet curvature

S. Myint et al. JPL



Large terrain modeling

Support storing and loading oflarge planetary scale data

Support data sets that cannot fitinto memory

Support random access

Be able to access the data in realtime

S. Myint et al. JPL



Paging

Only a subset of terrain datais paged in

SimScape builds on top ofHDF5 to support paging

In the rover example, only asmall patch of data underthe rover is kept in memory

S. Myint et al. JPL



Data tiling

Large data sets are sometimesbroken up into separate tiles

Mainly useful when modifying andwriting data (less memoryconsumption at any one time)

Useful for parallel processing ofdata (running on supercomputer)

Many data sets (e.g. MOLA) comein tiled format

Transparent to the data loadingAPI

S. Myint et al. JPL



Continuous level of detail visualization

A continuous level of detail(CLOD) technique allows us torender high resolution data onlywhere we want it (e.g, where thecamera is pointing).

Render very large data setsthat can’t normally berendered all at once by thegraphics card

Render in real time

Unlike a discrete LODtechnique, transitionbetween levels is smooth (nopopping between levels)

S. Myint et al. JPL



Continuous level of detail visualization example

S. Myint et al. JPL



Clipmapping

S. Myint et al. JPL



Clipmapping

Concentric rings of data

Innermost rings have thehighest resolution data

Each ring is composed of aregular grid

S. Myint et al. JPL



Clipmapping takes advantage of perspective projection

S. Myint et al. JPL



Other advantages of using clipmapping

Each ring has the samebasic regular grid geometry

Nearly the same operationsdone to each vertex

Computation easilyparallelizable (amenable toGPU implementation)

Full resolutionHalf resolution

Quarter resolution

S. Myint et al. JPL



GPU implementation

Implementing clipmapping on the GPU allows usto utilize hundreds of GPU cores.

Upload height map data to the GPU as atexture

Vertex program computes vertexpositions/normals based on:

Center position of all clipmapsClipmap ring levelHeight map texture

Fragment program computes final pixel colorbased on:

Albedo sampled from textureSurface normal (passed in from vertexshader)

Vertexprocessor

Rasterization

Fragmentprocessor

Frame buffer

S. Myint et al. JPL



Moving clipmaps

The high resolution area is movedby moving all clipmap rings.

S. Myint et al. JPL



Overlays

Wheel tracks areoverlaid on theterrain byperturbing thesurface normals ona per-pixel basis.

S. Myint et al. JPL



Overlays (continued)

Height maps

Albedo maps

Slope maps

S. Myint et al. JPL



Summary and future work

Summary

With paging, we can work with arbitrarily large terrain datasetsWe can visualize these large terrains using continuous level ofdetailWe make use of the GPU’s multi-core architecture in ourimplementation of continuous level of detail

Future work

Support paging of high-resolution (albedo) textures (instead ofjust geometry)Improve efficiency of GPU programs by dynamicallyautogenerating them based what features are being used (likeheight maps)

S. Myint et al. JPL


scalable large, multi-resolution terrain real-time...

Documents