thomas j. pingel & keith c. clarke department of geography university of california, santa...

19
DEMs for Immersive Geographic Virtual Environments: An Improved Simple Morphological Filter for Terrain Classification of LIDAR Data Thomas J. Pingel & Keith C. Clarke Department of Geography University of California, Santa Barbara AAG Annual Meeting, New York City, 24 Feb 2012

Upload: nedaa

Post on 22-Mar-2016

34 views

Category:

Documents


2 download

DESCRIPTION

DEMs for Immersive Geographic Virtual Environments: An Improved Simple Morphological Filter for Terrain Classification of LIDAR Data . Thomas J. Pingel & Keith C. Clarke Department of Geography University of California, Santa Barbara. AAG Annual Meeting, New York City, 24 Feb 2012. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Thomas J. Pingel & Keith C. Clarke Department of Geography University of California, Santa Barbara

DEMs for Immersive Geographic Virtual Environments: An Improved Simple Morphological Filter

for Terrain Classification of LIDAR Data

Thomas J. Pingel & Keith C. ClarkeDepartment of Geography

University of California, Santa Barbara

AAG Annual Meeting, New York City, 24 Feb 2012

Page 2: Thomas J. Pingel & Keith C. Clarke Department of Geography University of California, Santa Barbara

Project Overview

Build real-time geodatabases from audio and video feeds, and project them onto an immersive virtual world.

This immersive visualization is intended to aid in the understanding of a very recent or in-progress local event.

Page 3: Thomas J. Pingel & Keith C. Clarke Department of Geography University of California, Santa Barbara

The test bed:Isla Vista & the campus at UC Santa Barbara

Page 4: Thomas J. Pingel & Keith C. Clarke Department of Geography University of California, Santa Barbara

Good terrain layers are fundamental.

• Any errors will propagate through the rest of the VE construction process.

• Misshapen ground layers are confusing to the eye.

• A good ground layer can replace some kinds of extra information likely to be lacking.

Page 5: Thomas J. Pingel & Keith C. Clarke Department of Geography University of California, Santa Barbara

Requirements

• A good LIDAR-to-DEM production tool should be– Efficient with computation and memory– Validated against samples– Flexible

• Urban, suburban, and rural environments• Highly differentiated terrain

– Integrated• Specialized software is hard to validate• It lengthens the production chain, making automation difficult.

• A tool oriented to produce DEMs for visualization (instead of analysis) has particular issues as well.

Page 6: Thomas J. Pingel & Keith C. Clarke Department of Geography University of California, Santa Barbara

General Workflow Diagram

Identification of DSM cells as bare earth / object

Generate Digital Surface Model

Identify ground points from provisional DEM

Create provisional DEM

Page 7: Thomas J. Pingel & Keith C. Clarke Department of Geography University of California, Santa Barbara

open( I ) = dilate(erode( I ) )

I

erode( I )

open( I )

Morphological Opening

Page 8: Thomas J. Pingel & Keith C. Clarke Department of Geography University of California, Santa Barbara

Cross Section View of Image Opening

Page 9: Thomas J. Pingel & Keith C. Clarke Department of Geography University of California, Santa Barbara

When windowSize = [0 1 2 5 10 15], slope = 15% and elevationThreshold = .5

A sample progression of SMRF

Page 10: Thomas J. Pingel & Keith C. Clarke Department of Geography University of California, Santa Barbara

Other Notable Filters• Zhang et al. (2003)

– Exponentially increasing window size– Slope threshold based on difference in window sizes between steps

• Chen et al. (2007)– Applied a different method for vegetation and buildings– Object “prospects” were evaluated based on the distribution of

slopes around the perimeter• Other notable algorithms (not PMFs)

– Axelsson (1999) - Adaptive TIN– Shao (2007) – Climbing and Sliding– Meng et al. (2009) – Multidirectional

Page 11: Thomas J. Pingel & Keith C. Clarke Department of Geography University of California, Santa Barbara

Measuring Performance• ISPRS Datasets

– Sithole & Vosselman (2003 & 2004)– 15 samples in urban and rural environments– Less dense than most modern systems (.67 & .18 RPSM)

• Type I Error– BE as Object– causes “holes” in the DEM→ overly smooth areas

• Type II Error– Object as BE– causes overly rough areas

• Total Error & Cohen’s Kappa

Page 12: Thomas J. Pingel & Keith C. Clarke Department of Geography University of California, Santa Barbara

• [DTM groundIDs] = smrf(x,y,z,c,wk,s,[e1 e2])– c – cell size

• Related to resolution of input data – wk – maximum window size

• Vector of increasing values up to the size of the largest feature to be removed.

– s – slope threshold• Value of largest common terrain slope• Establishes elevation threshold for each step

– e – elevation threshold• Difference from digital terrain model (DTM) that is still identified

as ground.• Slope dependent threshold

Page 13: Thomas J. Pingel & Keith C. Clarke Department of Geography University of California, Santa Barbara

1) Create a copy of the DSM called lastSurface2) For thisWindow = 1 to maxWindow

a) thisThreshold = slope * (thisWindow / cellSize) b) thisSurface = open(lastSurface,disk(thisWindow))c) groundMask = groundMask OR

(lastSurface – thisSurface > thisThreshold)

d) lastSurface = thisSurface

Identification of DSM cells as bare earth / object

Page 14: Thomas J. Pingel & Keith C. Clarke Department of Geography University of California, Santa Barbara

SMRF vs. other PMFs• Oriented to reducing Type I error, while maintaining acceptable

Type II error rates• Built to be as simple as possible to provide a solid base from which

to test novel techniques• Linearly increasing window size, one-parameter based slope

thresholding• Uses PDE-based image inpainting instead of nearest neighbor /

kriging• Accepts a slope-based thresholding parameter for provisional DEM

to ground ID stage• Optional “net-cutting” routine to remove large buildings on

differentiated terrain.

Page 15: Thomas J. Pingel & Keith C. Clarke Department of Geography University of California, Santa Barbara

How well does SMRF perform?

• Single Parameter– Mean Total Error = 4.4%• Axelsson (4.82), Chen (7.23), Shao (4.20)

– Mean Kappa = 85.4%• Axelsson (84.19), Meng (79.93)

• Optimized– Mean Total Error = 2.97%– Mean Kappa = 90.02%

Page 16: Thomas J. Pingel & Keith C. Clarke Department of Geography University of California, Santa Barbara
Page 17: Thomas J. Pingel & Keith C. Clarke Department of Geography University of California, Santa Barbara
Page 18: Thomas J. Pingel & Keith C. Clarke Department of Geography University of California, Santa Barbara
Page 19: Thomas J. Pingel & Keith C. Clarke Department of Geography University of California, Santa Barbara

Future Work

• Public testing: search for LIDAR + SMRF online• Investigate more complex subroutines for

performance benefits• Data structures for VR display– Level of Detail, Grids / TINs

• Immersive DEM correction• Building reconstruction• True orthovideo overlay