15 March 2001 FINAL REPORT: An Integrated Feasibility Demonstration for Automated Population of Geos

1
15 March 2001 FINAL REPORTAn Integrated
Feasibility Demonstration for Automated
Population of Geospatial DatabasesSRI Project
Number 1515Contract Number
NMA202-97-C-1004DARPA Order Number
E645Prepared by Aaron Heller, Christopher
Connolly, Lynn Quam, Martin Fischler, Robert
BollesPrepared for Mr. Michael A.
OBrien National Imagery and Mapping
Agency 12300 Sunrise Valley Road Reston, VA
20191
Your logo here

• SRI International 333 Ravenswood Avneue Menlo
  Park, CA 94025

2
Introduction

• Program Goals
• Requirements
• Key Ideas
• Examples
• Conclusions

To see further discussion in the Lecture Notes of
this PowerPoint file, select View -gt Notes Page.
3
Main Program Goal

• Radical reduction in the cost, time, and human
  effort needed to model selected segments of the
  earths surface from remote-sensed data.
• Primary source is panchromatic/electro-optical
  imagery
• 1 to 2 orders of magnitude reduction
• days to hours

4
Other Program Goals

• Progress toward complete automation
• Assembly and community-wide availability of
  validated datasets and reference extractions of
  roads and buildings
• Metrics, procedures, and data formats for
  interchange, evaluation, and reporting of results

5
Requirements

• Remove the human from the processing pipeline
• Provide a systematic basis for achieving robust
  performance to avoid the need for significant
  human inspection and editing

6
Key Ideas

• Context-Based Algorithm Control System (CBACS)
• determine the relationship between contextual
  conditions and optimal algorithm settings
• integrated a 3-D world model to establish context
• The cost of an automated algorithm is the human
  effort required to verify the results and fix the
  errors
• computer time is essentially free

7
Architecture
Feature Extraction Managers
High-Level Task Descriptions
...
Compact 3-D Structures
Dynamic Objects
Terrain
Linear
Area
Application Service Modules
Context-Based Algorithm Control System (CBACS)
ATTRIBUTED FEATURE DATABASE
SEDRIS
VRML
Preprocessing/ Filtering and Registration
AND
MSE/ RADIUS
Feature Extraction Algorithm Suite
SELECTED RAW IMAGERY
IPT
IFSAR
...
SAR
EO
DTED
Existing Feature Data
8
Rural Road Extraction Assets

• Metadata
• Maps
• NIMA foundation data
• Region-based context (from imagery or HUMINT)
• Low-resolution road extraction
• High-resolution road refinement
• Semantic filters (e.g., road grade, curvatures,
  material type)
• Human editing

9
Types of Road

• Rural
• random pattern
• no adjacent correlated structure
• Suburban
• Semiregular pattern
• houses, driveways, parked cars
• relatively unoccluded
• Urban
• regular pattern
• heavily occluded

10
Road Network Modeling Data Flow
11
Rural Road Extraction Strategy

• Interactively outline rural, suburban, and urban
  areas
• Apply low-resolution analysis to rural areas with
  loose parameters to generate a superset of the
  valid road segments (using a priori map, if
  available)
• Apply high-resolution refinement process
• Use semantic filters to reclassify nonroads
• Suggest new segments to link isolated pieces to
  nearby segments (including a priori map segments)
• Reapply low- and high-resolution analysis to
  examine suggested links
• Provide robust capability to support rapid,
  interactive edit of final results

12
Ft. Benning McKenna MOUT
13
Low-Resolution Analysis
14
Low-Resolution Analysis
15
Road Network
16
High-Resolution 3-D Analysis
17
Spur Detection
18
Final Result
19
Evaluation Reference Model
20
Evaluation Scoring
21
Results to Date

• Ft. Benning McKenna MOUT
• 3 min operator, 45 min CPU (GDE ref 270 min)
• multiple vertical panchromatic images
• 6 km2 area 20 km roads extracted
• 99 completeness and correctness (GDE ref model)
• Ft. Irwin/NTC built-up area
• 10 min operator, 230 min CPU (NIMA ref 400 min)
• DPPDB (NTM)
• 20 km2 area 65 km roads extracted
• 95 completeness and correctness (NIMA DTOP ref
  model)
• NTC full-frame NTM stereo pair
• OCONUS site

22
Ft. Irwin/NTC Built-Up Area
23
A New Site - Ft. Stewart, Georgia

• Description
• Visually similar to Ft. Benning/McKenna site
• 300 km2, covered by six USGS digital ortho quads
  (DOQ) (1m ground sample distance (GSD))
• USGS 30m DEM
• Experiment
• Does automatic road extraction system perform
  well on this site with out of the box settings?
• Results
• Ran low-level detector at 2, 4, and 8m GSD
  merged results
• Ran high-level processing at full resolution

24
Automated Result
25
Kennedy Space Center
26
Automated Result
27
Building Extraction Architecture

• Generate cue points or footprints
• interactively
• IFSAR or dense DEM analysis
• Extract geometry
• bottom-up (USC), top-down (GDE)
• Refine geometry
• multi-image model-based optimization (SRI)
• Review and edit

28
Cue Points from ISFAR and MS
29
Result
30
Building Footprints from Dense DEM
31
Building Footprints from Dense DEM (2)
32
Buildings from Dense DEM
33
Other Accomplishments

• APGD virtual lab
• Data sets and reference models
• SEDRIS interface
• Sensor model API
• Import procedures for NIMA products
• CIB, DPPDB, VPF, DTED
• RCDE/SocetSet interoperation
• Algorithms for LULC and IFSAR analysis

34
Conclusions

• State-of-the-art performance
• Fully automatic techniques provide a good
  approximation to desired feature extraction
• Prototype IU-assisted editors bring result up to
  product standards
• The APGD program has the goal of advancing
  extraction of selected cartographic features by 1
  to 2 orders of magnitude. We feel that we have
  been successful in demonstrating the feasibility
  of this goal.

35
(No Transcript)