PowerPoint template tips

1
(No Transcript)
2
Automatic Linear Feature Identification and
Extraction(ALFIE)

• Richard Ley

3
Contents of presentation

• 1 Introduction
• 2 Innovation
• 3 System architecture
• 4 Achievements
• 5 Exploitation

4
Introduction

• Section 1

5
The ALFIE team

• QinetiQ
• Project Manager
• Technical Lead
• Team
• Laser-Scan Limited
• Nottingham University (School of Geography)

6
TG5 R06 Desired Outcomes

• D04 Tools and techniques to improve the fidelity
  and realism with which the real world can be
  modelled
• D05 Development and implementation of
  standards/architectures to support economy of
  SE provision
• D07 Tools and techniques to improve efficiency
  (time and cost) of preparation,integration,
  operation and evaluation of SEs

7
The capability gap (1)

• Rapid and cost effective Synthetic Natural
  Environment (SNE) terrain database generation
• new terrain databases required to support all
  aspects of training, in particular mission
  planning, mission rehearsal
• database generation is time consuming and
  expensive
• essential source data includes digital
  information products such as DTED and VMap
• digital products are often out of date,
  inaccurate and resolution inadequate

8
The capability gap (2)

• requirement for accurate, up-to-date, worldwide
  coverage of geospatial information on demand
• remotely sensed imagery provides a source of
  accurate and up-to-date data
• automated processes are available to extract
  elevation information from remote sensed imagery
• automated processes for feature extraction do not
  exist
• automated feature extraction and attribution
  essential for rapid/cost effective generation of
  SEs

9
Future goal architecture
Graphical User Interface
Raw Raster Data
Raw Vector Data
Networked Texture Libraries
Standard Interchange Format
Database
3D Models Libraries
Standard Products
Image Processing Software Component
Softcopy Photogrammetry Software Component
Additional Data
10
ALFIE goals

• To develop a proof of principle capability to
  demonstrate automatic extraction of linear
  feature information
• roads (dual carriageways and other roads),
  railways and rivers
• presentation will include screen shots from
  demonstration
• To support the production of accurate, up-to-date
  geospatial information on demand for
• production of maps, charts and databases for
  mission planning, mission rehearsal, operations,
  training simulations, etc

11
Innovation

• Section 2

12
Semi-Automated approach

• Manual selection of imagery
• Manual identification of individual features
• Semi-Automatic extraction
• Semi-Automatic topological construction
• Manual validation
• Manual post-editing
• Automation of extraction reduces timescales but
  still limited to algorithms tailored to specific
  image types

Extraction
Network Build
Post editing
13
ALFIE approach

• Automatic selection of imagery
• Automatic linear extraction
• Automatic identification (classification)
• Automatic topological construction
• Automatic validation
• Manual post-editing
• Use of a control strategy renders the system
  automatic
• reduced timescales reduced cost rapid terrain
  database generation

14
ALFIE philosophy

• Develop a framework that facilitates automatic
  extraction, identification and attribution of
  linear features
• Build upon state of the art
• Multi-resolution imagery
• Concept of minimal and optimal datasets
• Tool-kit of algorithms
• Exploit contextual information
• Spatial and aspatial
• Develop control strategy
• Feature oriented solution (Object oriented
  database)

15
Process flow
Preparation
Pre - Processing
Collateral Extraction
Extraction
Classification
Validation
16
System Architecture

• Section 2

17
Preparation (1)

• Goal
• Selection of imagery
• Automatic selection of algorithms
• Requirement
• Different image types require different
  extraction algorithms
• Automatic selection of most appropriate algorithm
  required
• Automatic selection of most appropriate
  parameters

18
Preparation (2)
Effect of differences in spectral and spatial
resolution
Landsat Image (7 bands - 30m resolution)
KVR Image (1 bands - 2m resolution)
19
Preparation (3)
20
Pre-processing (1)

• Goal
• Prepare selected image for operation of selected
  feature extraction algorithm
• Requirement
• Operation of algorithms can be optimised by
  careful selection of pre-processing techniques
• e.g. smoothing to reduce noise, or edge
  enhancement to increase effectiveness of edge
  detector

21
Pre-processing (2)
22
Collateral extraction (1)

• Goal
• Extract information required for the
  classification of extracted linears and
  associated processes
• Requirement
• Extraction of masks which enable higher
  classification confidence e.g. water
• Derivation of information required to build and
  complete networks e.g. junctions

23
Collateral extraction (2)
NDVI Highlights vegetation
Texture Indicates homogeneous areas
24
Collateral extraction (3)
Preparation
Pre - Processing
Collateral Extraction
Extraction
Original Image
Water Mask
Classification
Initial classification of linears can be refined
by intersecting these with the water mask to
provide a higher confidence classification
Validation
25
Collateral extraction (4)
VMap data
26
Linear extraction (1)

• Goal
• Extract features and populate OO database with
  unknown class
• Clean and build lines
• Requirement
• Extract centrelines or edges
• Process and clean lines
• Local snapping to work with longer linears
• Mask urban/rural areas

27
Linear extraction (2)
28
Linear extraction (3)
29
Linear classification (1)

• Goal
• Determine probability of each extracted linear
  falling into each feature class
• Iterate until highest confidence of class
  membership found
• Requirements
• Generate evidence of class membership
  (discriminants)
• Assess evidence of class membership - Cluster
  weighted model (generated under TG10) - dual
  carriageways, other roads, railways and rivers

30
Linear classification (2)Discriminants

• Geometric
• Length, curvature, angle of turn, sinuosity,
  orientation
• Geometric/photometric
• Width, variation in width
• Photometric
• Dominant spectral value, mean spectral value,
  variation of spectral value, significant
  discontinuities
• Elevation
• Dominant elevation, mean elevation, gradient of
  elevation, significant discontinuities

31
Linear classification (2)Discriminants

• Geometric
• Length, curvature, angle of turn, sinuosity,
  orientation
• Geometric/photometric
• Width, variation in width
• Photometric
• Dominant spectral value, mean spectral value,
  variation of spectral value, significant
  discontinuities
• Elevation
• Dominant elevation, mean elevation, gradient of
  elevation, significant discontinuities

32
Linear classification (3)Initial classification
result
Initial Extraction Result
33
Linear classification (4)Network building
Find candidate junctions
34
Validation (1)

• Goal
• Assess extraction and attribution accuracy and
  completeness of final linear network structure
• Requirement
• No automated extraction will be 100 correct.
  Methods are required to assess graphically and
  statistically the accuracy of the result
• Provide an intuitive and straightforward means of
  editing the results

35
Validation (2)Classification confidence
Classification confidence
36
Validation (3)Connectivity
Fully connected
Hanging endpoints
Isolated linears
37
Validation (4)Classification performance
38
Validation (5)Classification performance
39
Achievements

• Section 4

40
Achievements

• Conceptual level
• Implementation level

41
Conceptual level (1)

• Modular system provides future proofing
• technology insertion
• Open and scalable architecture
• COTS IP and GIS technologies
• Object Oriented data model
• Control Strategy
• flexible output-driven
• automates entire process

42
Conceptual level (2)

• Extraction
• Delineation
• Dumb
• Classification
• Context based attribution
• geometric photometric
• Network building
• Identification

43
Conceptual Level (3)

• Automated the entire process
• International recognition of approach
• I/ITSEC
• ISES
• Ascona
• ISPRS

44
Future goal architecture
Graphical User Interface
Raw Raster Data
Raw Vector Data
Networked Texture Libraries
Standard Interchange Format
Database
3D Models Libraries
Standard Products
Image Processing Software Component
Softcopy Photogrammetry Software Component
Additional Data
45
Implementation Level

• Specific solution
• built on Laser-Scans Gothic OODB
• public domain and QinetiQ algorithms
• tool kit
• standard hardware and software

46
Implementation Level - Results (1)
Control Strategy

• Control interface
• invokes appropriate control modules
• provides flexible automatic process flow
• but allows user interaction

47
Implementation Level - Results (2)

• automatic selection of imagery and algorithms
  against requirement
• automatic optimisation of the imagery
• automatic extraction and attribution
• 62 achieved urban region
• 96 achieved rural region
• automatic classification
• cluster weighted model
• 75 achieved test dataset
• 63 achieved extracted dataset
• network building

Preparation
Pre - Processing
Collateral Extraction
Extraction
Classification
Validation
48
Exploitation

• Section 5

49
Exploitation

• MoD Pull-through
• Applied Research
• ARP 7 Combat Readiness, Training Simulation
• ARP 13 Joint TOC
• ARP 14 Imagery Exploitation
• Integrated Project Teams
• JTBSE
• Images

50
Exploitation

• Work identified in STOW programme
• capability gap
• rapid terrain database generation

51
Exploitation

• Work identified in STOW programme
• capability gap
• rapid terrain database generation
• Next stage
• address 3-D
• ALFIE-2
• SE, SS CRP Project

Softcopy Photogrammetric Software Component
52
The ALFIE-2 team

• QinetiQ
• Project Manager
• Technical Lead
• Team
• Laser-Scan Limited
• University College London (Department of Geomatic
  Engineering)

53
ALFIE-2 architecture
54
Building Extraction (1)

• Existing automated techniques for extracting
  terrain surface are limited at extracting
  features
• Extracted terrain is typically smoothed and
  generalised
• Stereomatching can be improved by using context
• knowledge of breaklines, e.g. building edges
• this information can be derived automatically by
  the ALFIE process
• Positive feedback loop
• attributed linears aids feature matching
• 3-D information as an additional discriminant

55
Building Extraction (2)
Extraction without features
Extraction with features
56
Conclusions

• Motivated and well integrated team
• Proven approach
• flexible, extensible, future proofed
• innovative, extraction and attribution of linear
  features
• Fully automated proof of principle system
  developed
• based on COTS software
• Need for better discrimination between
  roads/railways
• use of complementary datasets e.g. SAR
• Need to improve initial feature extraction
• Need for high accuracy 3-D information
• ALFIE-2

57
Contact Details Brenda Stroud Project
Manager QinetiQ (Fort Halstead) bjstroud_at_Qin
etiQ.com Richard Ley Technical Lead QinetiQ
(Farnborough) rgley_at_QinetiQ.com