Comprehensive Survey of Extraction Techniques of Linear Features from Remote Sensing Imagery for Upd

1
Comprehensive Survey of Extraction Techniques of
Linear Features from Remote Sensing Imagery for
Updating Road Spatial Databases
Ji Sang Park, PhD Candidate Dr. Raad A. Saleh,
Scientist

• Department of Civil and Environmental Engineering
• University of Wisconsin-Madison
• ERSC 12th Floor
• 1225 West Dayton Street, Madison, WI 53706
• Phone [phone]
• Fax [phone]

2
Comprehensive Survey of Extraction Techniques of
Linear Features from Remote Sensing Imagery for
Updating Road Spatial Databases
Abstract

• Research on automated and semi-automated
  extraction techniques of linear features from
  remote sensing imagery has been active for
  decades. Features of interest include
  transportation networks, power transmission
  lines, etc. 
• This paper presents a comprehensive survey of
  extraction techniques of such features from
  aerial and satellite imagery.  The techniques are
  evaluated with respect to methodology, strengths,
  drawbacks, and implementation approach.  Source
  data for the surveyed techniques include
  panchromatic and multispectral imagery.  The
  viability of hyperspectral data is extrapolated
  for same purpose of utilization. The paper later
  presents a discussion of automated extraction
  techniques specifically used for updating road
  spatial databases. 

3
Outlines

• GIS Data of Roads
• Characteristics of Roads
• Problems in Extracting Roads from Imagery
• Road Detection Methods
• Road Tracking Methods
• Trends

4
GIS data of Roads

• National Highway Planning Network
• BTS data
• Federal Highway Administration
• Scale 1100,000
• Representing 400,000 miles of federal roads in 50
  states including Puerto Rico
• DB updating method varies

5
(No Transcript)
6
GIS data of Roads

• GDT
• Geographic Data Technology, Inc.
• Enhanced TIGER DB
• Using DOQ and satellite imagery to update their
  spatial DB

7
(No Transcript)
8
GIS data of Roads

• NAVTECH
• Navigation Technology, Inc.
• Using existing maps
• Digitizing based on aerial photographs
• Driving and testing with GPS

9
(No Transcript)
10
Characteristics of Roads

• Radiometric
• Various grayscale along road extent
• Relatively constant grayscale and texture between
  boundaries
• Spectral
• Consistent signature
• Spectral response depends on material

11
Characteristics of Roads

• Geometric
• Long and continuous
• Narrow width
• Two-lane 4.8m(16ft) 7.2m(24ft) with 3m
  shoulders
• Divided 3.6m(12ft) travel lane with 6m(20ft)
  wide median strip
• With small curvatures
• Different shapes
• High-resolution Rectangular objects with
  parallel boundaries
• Low-resolution Linear objects

12
(No Transcript)
13
(No Transcript)
14
Problems in Extracting Roads from Imagery

• Radiometric
• Line disconnection due to covering over roads
• Trees, shadows, buildings, and vehicles
• Detection of wrong objects or areas due to
  similar grayscale
• Objects or surrounding of roads
• Blurred boundaries of roads
• Spectral
• Different spectral information due to camera
  angle, atmospheric distortions, etc.
• Inconsistent spectral response
• Inaccurate signatures

15
Problems in Extracting Roads from Imagery

• Geometric
• Different horizontal profiles due to various
  widths and types of roads
• Number of lanes
• Divided / undivided
• Short or dead-end road
• Note
• Important to keep balance between detection
  performance and local condition.
• The more edges are extracted, the more complex
  they become.

16
Road Detection Methods

• Using radiometric information
• Using geometric information
• Using LIDAR data

17
Using radiometric information

• Convolution or image segmentation.
• Popular method for approximating initial road
  regions.
• Amount of data is reduced significantly while
  retaining structural information of features of
  interest.
• Most of the methods adopt Gaussian smoothing to
  reduce small details.

18
Using radiometric information

• Methods
• Convolution
• High pass filter detect high frequency
• Canny filter global position of tracked
  discontinuities
• Nevatia-Babu filter edge detection edge
  thinning
• Gradient Direction Profile Analysis (GDPA)
• Determine gradient direction for a pixel as the
  direction of maximum slope.
• Image segmentation
• Watershed transform
• Partitions an image into homogeneous regions.
• Locates gradient contours based on the gradient
  magnitude and direction.
• Assisted by multiscale image analysis
• Indicate global location and relative size of
  terrain objects.

19
Using radiometric information

• Methods
• Signal processing
• a trous algorithm
• Multiresolution analysis (MRA)
• Eliminate small particles by smoothing
• Describe the hierarchical information of
  features.
• Wavelet transform
• Establish a local relationship between a spatial
  domain and a frequency domain.
• Approximate the first derivatives of the pixel.
• Computation of successive approximations by
  smoothing.
• Determine edges based on wavelet coefficients.
• Neural Network
• Dynamic programming
• Defining a cost which depends on local
  information
• Summation minimization process

20
Using radiometric information

• Convolve the image in the spatial domain using an
  appropriate kernel
• Kernels can be used for connecting segments
• Connected components are labeled

21
Using geometric information

• Methods
• Convolution
• Direction filter direction of extracted regions
• Parallel edge detection filter parallelism of
  edges
• Optimal search algorithm
• Distances and directions between road segments
• Hough transform
• Connectivity of line segments can be computed
  analytically
• Tolerant of gaps in feature boundary descriptions
• Using templates and models

22
Using LIDAR data

• Berg. R. and Ferguson, J. (2000)
• Classification was primarily for removal of
  vegetation data
• Where applicable, building data were also removed
• Possible for road shaping and line linking
• Rigorous manual analysis and edit was required

23
Using LIDAR data

• Photogrammetric mapping provides a better
  representation of narrow features since accurate
  breakline data points can be collected directly
  along the feature of interest
• Not effective for feature mapping
• The raw data points may not be located directly
  on the features.
• Does not define breaklines along features.
• Advantages
• Density of points
• Ability to penetrate canopy
• Effective for large project area within short
  time period

24
Road Tracking Methods

• Hough Transform
• Optimal Search
• Profile Analysis

25
Hough Transform

• Computing global description of features with
  given measurements.
• Determine both WHAT the features are and HOW MANY
  of them exist.
• Parametric description of a line

y
xcos? ysin? r
(x, y) -gt (r, ? )
r
?
x
26
Hough transform

• Procedures
• Points in cartesian image space map to curves in
  the polar Hough parameter space.
• Curves by collinear points intersect in peaks (r,
  ? ).
• Intersection points characterize the straight
  line segments.
• Extract local maxima from the accumulator array
  (relative threshold).
• Mapping back from Hough space into image space.

27
Hough transform

• Advantages
• Tolerant of gaps in feature boundary.
• Unaffected by image noise.
• Disadvantages
• Distance between points on lines is not
  considered.

28
Optimal search

• Directional cone search (Lee et al. 1999)
• Represent local trend of features
• Searching process
• Shoot two cones with the direction of the region.
• The cones may meet several regions.
• Choose the most probable road region and connect
  two regions by adding regions between two
  regions.
• Repeat from the beginning until no more
  reconnection occurs.

29
Optimal search

• Directional cone search
• Useful when roads are defined as long rectangular
  objects.
• Tracking result is good in urban area.
• Affected by image noises.

30
Profile analysis

• GDPA (Gradient Direction Profile Analysis)
• Gradient direction direction of max slope among
  four defined directions near a pixel
• A1 a4 g a8 g / 2
• Perpendicular to the ridge for the pixel
• Highest point correspond to the top of ridge.
• Additional fitting function is used between steep
  slopes and gentle slopes.

A2
A3
A4
A1
31
Profile analysis

• GDPA
• Advantages
• Edge detection road tracking are done
  simultaneously.
• Describe local conditions of features.
• Simple procedure using only gradient value.
• Disadvantages
• Similar radiometric contrast between roads and
  surroundings provides bad result.
• By using small size convolution window, tracking
  effect is not good in urban area due to complex
  structures and various obstacles.

32
Trends

• Strategies
• Using both radiometric and geometric information
• Radiometric find road regions in images
• Geometric construct parallel boundaries and link
  disconnected road segments
• Image resolution
• High-resolution matching profile and detection
  of road sides
• Low-resolution detection and following of lines

33
Trends

• Strategies
• Exploiting GIS layer
• Can be used for road linking, but not for road
  positioning
• Using LIDAR data
• Can be used for road shaping and linking as a
  reference data

34
Possible operators for road detection
 
35
Trend
Input Images
Road Region Finding
Canny Filter
Parallel Edge Filter
Road Shaping
LIDAR
Hough / Optimal
Road Linking
GIS Layer (Optional)
Thinning / Vectorizing
SOM, Snakes
Road Network
36
SAR Imagery

• A SAR SPECKLE FILTERING ALGORITHM TOWARDS EDGE
  SHARPENINGYunhan Dong, Anthony K Milne, and
  Bruce C ForsterSchool of Geomatic Engineering,
  Office of Postgraduate StudiesThe University
  of New South WalesSydney 2052,
  Australiay.dong_at_unsw.edu.au, t.milne_at_unsw.edu.au,
  b.forster_at_unsw.edu.auWorking Group VII/6

37
Filters Applied on Non-Edge Features
38
Filters Applied on Edge Features
39
GIS Assisted Feature Extraction

• MATCHING LINEAR FEATURES FROM SATELLITE IMAGES
  WITH SMALL-SCALE GIS DATA
• Reference
• Andreas BUSCH
• Bundesamt fur Kartographie und Geodasie
• Richard-Strauss-Allee 11
• 60598 Frankfurt am Main, Germany
• busch_at_ifag.de

40
GIS-Image Analysis Flow
GIS
Prior Information
Revision
Image Analysis
Flow of information between GIS and image
analysis.
41
Measures and Criteria for Matching

• All possible correspondences within the
  neighborhood defined by a maximal distance there
  is need for measures to evaluate the quality of
  different matches.
• Distance
• Length
• Parallelism
• Semantics

42
(No Transcript)
43
(No Transcript)
44
INTEGRATED GEOGRAPHIC INFORMATION SYSTEMS IMAGE
ANALYSIS
INTEGRATED GEOGRAPHIC INFORMATION SYSTEMS FROM
DATA INTEGRATION TO INTEGRATED ANALYSIS Reference
Manfred EHLERSInstitute for Environmental
SciencesUniversity of VechtaP.O. Box 1553,
D-49364 Vechta, Germanymehlers_at_ispa.uni-vechta.de
45
(No Transcript)
46
CARTOGRAPHIC FEATURES FROM AERIAL IMAGES
AUTOMATIC CARTOGRAPHY FROM AERIAL IMAGES (SITE OF
SALE, MOROCCO) Reference O.El
Kharki,M.Sadgal,A.Ait Ouahman,A.El
Himdy,M.Ait BelaidLaboratory of Electronic
and Instrumentation,Faculty of Science Se
lalia,BOX 2390 Marrakech,Morocco.elkharki_at_yahoo.f
rAd inistration de la Conservation Foncière du
Cadastre et de la Cartographie,Rabat,Morocco.R
oyal Centre for Remote Sensing,Rabat,Morocco.
47
(No Transcript)
48
METHODOLOGY

• Split and Merge Algorithm
• The basic idea of region splitting is to break
  the image into a set of disjoint regions which
  are coherent within the selves
• -Initially take the image as a whole to be the
  area of interest.
• -Look at the area of interest and decide if all
  pixels contained in the region satisfy some
  similarity constraint.
• -If TRUE then the area of interest corresponds to
  a region in the image.
• -If FALSE split the area of interest (usually
  into four equal sub-areas)and consider each of
  sub-areas as the area of interest in turn.
• -This process continues until no further
  splitting occurs.In the worst case this happens
  when the areas are just one pixel in size.

49
METHODOLOGY

• If only a splitting schedule is used then the
  final segmentation would probably contain any
  neighboring regions that have identical or
  similar properties.
• Thus,a merging process is used after each split
  which co pares adjacent regions and merges the if
  necessary.

50
(No Transcript)
51
SYSTEM TRENDS FOR AUTOMATED FEATURE EXTRACTION

• DIGITAL SYSTEMS FOR AUTOMATED CARTOGRAPHIC
  FEATURE EXTRACTION
• Reference
• Eberhard Gü lchUniversity of Bonn,
  GermanyInstitute of Photogrammetryebs_at_ipb.uni-bo
  nn.de

52
SYSTEM TRENDS FOR AUTOMATED FEATURE EXTRACTION

• 1. Interactive system (purely manual measurement,
  no automation for any measurement task)
• 2. Semi-automatic system (interactive environment
  and integration of automatic modules in the
  workflow)
• 3. Automated system (interactive environment with
  interaction before and after the automatic phase)
• 4. Autonomous system.

53
SYSTEM TRENDS FOR AUTOMATED FEATURE EXTRACTION

• Commercially available photogrammetric systems
  now include feature collection module
• In last year s comprehensive evaluation by GIM
  International (Plugers, 1999), there are 19
  digital photogrammetric workstations listed
• The basic input are digital or digitized images
  with the emphasis on stereo-imagery.

54
SYSTEM TRENDS FOR AUTOMATED FEATURE EXTRACTION

• Three types of methods are distinguished in the
  GIM survey
• Semi-automatic line extraction (7 systems)
• Semi-automatic corner point extraction (5
  systems)
• Automatic break-line extraction (3 systems)

55
SCALE-SPACE EXTRACTION TECHNIQUES

• MULTI-SCALE ROAD EXTRACTION USING LOCAL AND
  GLOBAL GROUPING CRITERIA
• Reference
• Albert Baumgartner, Stefan Hinz
• Chair for Photogrammetry and Remote
  SensingTechnische Universit at M unchen,
  D80290 Munich, GermanyE-mail
  falbertgfhinzg_at_photo.verm.tu-muenchen.deURL
  http//www.photo.verm.tu-muenchen.de

56
(a) Image (b) Segmentation of open rural context
57
(a) Initial hypotheses for road segments (b)
Detail
58
Results of local grouping.
Results of global grouping.
59
Results of integrated combination of local and
global module
Results of sequential combination of local and
global grouping.
60
(No Transcript)
61
(No Transcript)