Large-Scale 3D Terrain Modeling. David L. Page Mongi A. Abidi, Andreas F. Koschan Sophie Voisin, Sreenivas Rangan, Brad Grinstead, Wei Hao, Muharrem Mercimek Imaging, Robotics, & Intelligent Systems Laboratory The University of Tennessee March 23, 2004. Outline. 3D Terrain Modeling
Large-Scale 3D Terrain Modeling
David L. Page
Mongi A. Abidi, Andreas F. Koschan
Sophie Voisin, Sreenivas Rangan, Brad Grinstead, Wei Hao, Muharrem Mercimek
Imaging, Robotics, & Intelligent Systems Laboratory
The University of Tennessee
March 23, 2004
Multi-sensor data collection system for road surface.
GPS Base Station
3-Axis IMU and Computer
3D Range Sensor
1 – Riegl LMS-Z210 Laser Range Scanner
2 – SICK LMS 220 LaserRange Scanner
3 – JVC GR-HD1 High Definition Camcorder
4 – Leica GPS500 D-RTK
Global Positioning System
5 – XSens MT9 Inertial Measurement Unit
6 – CPU for acquiring SICK, GPS, and IMU data
7 – CPU for acquiring Riegl data
8 – Power system
Mounted here on a push cart.
Geo-referenced geometric 3D model of an area near IRIS West in Knoxville.
3D View of Terrain
(Jump to 3D Viewer)
Why needed, in general?
Discussions with Dr. Al Reid
Scanning 3D terrains is a significant enhancement over traditional towed-cart profiling, cart dynamics, 1D profile, etc.
Real terrain modeling overcomes potential limitations of linearity, stationarity, and normality assumptions, particularly associated with PSD (Chaika & Gorsich 2004).
Research in 3D processing (tools!) addresses relevant issues in…
data reduction (Al Reid),
terrain analysis (3D EMD),
Towed Trailer Profilometer
Highly correlated sensor data (GPS, IMU, Range) = Correction for vehicle dynamics
UTK 3D Terrain Modeling
120-360 profiles over a 2-8 m swath (3D surface) vs. 1 profile (1D signal)
Correlated data vs. trailer dynamics
Agile path vs. linear path (?)
3D vs. 1D
Path Overlaid on Aerial View
2 m wide x 8 m length
Path is 300 m length +/- 0.5 cm resolution
Video Data of Zoom
Notice Cracks in Pavement
3D range sensors
Position and orientation sensors
3D Position and Orientation
Genex 3D CAM
IVP RANGER SC-386
Sheet-of-light triangulation-based system
Structured-light stereo system
Principle of operation
S1 and S2 are two sensors.
Standard Deviation = 0.0336
Standard Deviation = 0.0338
Standard Deviation = 0.0492
Extensive GPS and IMU error characterization and modeling.
Visualization tool built to be able to visualize “z” measurements
Blue Line is the GPS Path for the loops that we collected.
Cornerstone Drive, off Lovell Road, I-40 Exit #374 Knoxville, Tennessee, Knoxville
Each loop a length of 1.1 mile, Total distance covered on scanning that day = 2.2 miles ( 2 times) = 4.4 miles of the same data.
The color of the GPS path encodes the height of the terrain.
Over 4 miles = ~2 GB of data
Automated correction for varying speeds and dynamics of platform.
Pathways – Loop scanning
Full length scanning
The entire stretch,
Latitude and Longitude
from the Leica DGPS
Raw Point Cloud
Vehicle (Scanning) Direction
GPS curve sampled at 10 Hz.
IMU data @ 100 Hz
Video recorded at 30 frames/sec
Range Profiles @
30 Hz 4m wide SICK
2000 Hz and 50cms wide IVP
Correct for non-uniform data collection with terrain modeling.
Pose From Motion
Oriented Tracks Filtering
GPS drop-outs under certain conditions.
Improve overall localization accuracy.
Multiresolution Analysis and Denoising
Dataset from near IRIS West
The total length of the patch: 20 meters with inter-profile spcaing around 1 cm.
Reconstructed 3D profile from the statistical model
Mean Longitudinal profile
The 3D terrain was generated using our system mounted on a van.
The profile is non-linear and non-stationary but all the IMF’s taken separately are linear and stationary, which means the PSD of the IMF’s model the data better than the PSD of the profile alone.
Empirical mode decomposition of the terrain sample shown above.
EMD implementation : Modified Brad’s functions
Pipeline of 3D Reconstruction
(+/- 0.5 mm)
Tire 150 cm dia., 30 cm width
(18 Sections, 7 Views)
Pathways – Loop scanning