SLIDE 1 Automatic Extraction of Road Intersection Position, Connectivity, and Orientations from Raster Maps
Yao-Yi Chiang and Craig Knoblock
University of Southern California Department of Computer Science and Information Sciences Institute
SLIDE 2 Introduction
- Raster maps are one important source of
geospatial data:
- Contain information that is difficult to find elsewhere
- Contain the most complete set of data
1
USGS topographic map Map of Tehran, Iran
SLIDE 3 Introduction
- In [Chen et al. 2008], we utilize the set of road intersection
templates as the fingerprint of the raster map to integrate raster maps with other geospatial data
- Road intersection template:
- Road intersection position, connectivity, and road orientation
2
Satellite Imagery Raster map Vector data
SLIDE 4
USGS Topographic Map, El Segundo, CA USA
SLIDE 5
USGS Topographic Map, El Segundo, CA USA
SLIDE 6
USGS Topographic Map, El Segundo, CA USA
0∘
SLIDE 7
USGS Topographic Map, El Segundo, CA USA
0∘ 90∘
SLIDE 8
USGS Topographic Map, El Segundo, CA USA
0∘ 90∘ 180∘
SLIDE 9
TIGER/Line Vector Data with Geo-coordinate Information
SLIDE 10
TIGER/Line Vector Data with Geo-coordinate Information
SLIDE 11
TIGER/Line Vector Data with Geo-coordinate Information
SLIDE 12
TIGER/Line Vector Data with Geo-coordinate Information
SLIDE 13
TIGER/Line Vector Data with Geo-coordinate Information
SLIDE 14
TIGER/Line Vector Data with Geo-coordinate Information
SLIDE 15 TIGER/Line Vector Data with Geo-coordinate Information
Found the map location!!
SLIDE 16 Previous Work
- The accuracy of the road intersection templates is
important
- Help to prune the searching space during the matching
- Challenges for extracting the road intersection templates:
- Limited access to the metadata of the maps
- Maps are complex
SLIDE 17 Previous Work
- In [Chiang et al. 2008], we worked on the pixel level to
decompose the raster maps and to extract the road intersections automatically
Remove Background Remove Noise and Rebuild Road Layer Identify Road Intersection Candidates
SLIDE 18 Previous Work
- A simpler method to identify road intersections and extract
the road intersection templates
- We also determine the road format (i.e., single or double
line) and extract the road width
17
SLIDE 19 Previous Work
- A simpler method to identify road intersections and extract
the road intersection templates
- We also determine the road format (i.e., single or double
line) and extract the road width
18
Road Intersection!! Road width
SLIDE 20 Previous Work
- Lines are distorted by the thinning operator
- The extracted road intersection templates are not
accurate
19
SLIDE 21 Extract Accurate Road Intersection Templates
- The distortion is caused by using the thinning
- perator on thick lines
- The extent of the distortion is determined by the
road width
20
Rebuild road layer
SLIDE 22 Extract Accurate Road Intersection Templates
- In this work, we skip the distorted areas and trace
the straight lines to extract accurate road intersection templates
21
Road width
SLIDE 23 Extract Accurate Road Intersection Templates
22
The blob image The thinned lines Intersection Positions
SLIDE 24 Extract Accurate Road Intersection Templates
23
The blob image The thinned lines Intersection Positions The size of a blob is determined using the road with for covering the distorted lines
SLIDE 25 Extract Accurate Road Intersection Templates
24
The blob image The thinned lines Intersect the thinned line image with the blob image Intersection Positions The size of a blob is determined using the road with for covering the distorted lines
SLIDE 26 Extract Accurate Road Intersection Templates
- Identify contact points
- Trace road line candidates from contact points
25
SLIDE 27 Extract Accurate Road Intersection Templates
- Trace road line candidates from contact points
- Trace only a certain amount of line pixels to prevent looping
- Road lines are straight within a small distance (e.g., 5 pixels)
- Fit a line function (Y= aX+b) to the traced pixels using Least-
Squares Fitting algorithm
26
SLIDE 28 Extract Accurate Road Intersection Templates
- Trace road line candidates from contact points
- Trace only a certain amount of line pixels to prevent looping
- Road lines are straight within a small distance (e.g., 5 pixels)
- Fit a line function (Y= aX+b) to the traced pixels using Least-
Squares Fitting algorithm
27
SLIDE 29 Extract Accurate Road Intersection Templates
- Trace road line candidates from contact points
- Trace only a certain amount of line pixels to prevent looping
- Road lines are straight within a small distance (e.g., 5 pixels)
- Fit a line function (Y= aX+b) to the traced pixels using Least-
Squares Fitting algorithm
28
SLIDE 30 Extract Accurate Road Intersection Templates
- Update road intersection templates
- Keep every road line candidate
- Use the intersection of the line candidates to update the template
29
1 2 3 3 1 2 Original map Blob image intersected with thinned lines Traced line functions
SLIDE 31 Extract Accurate Road Intersection Templates
- Update road intersection templates
- Keep every road line candidate
- Use the centroid of the intersections of the line candidates to
update the template
30
1 2 4 3 Original map Blob image intersected with thinned lines
SLIDE 32 Extract Accurate Road Intersection Templates
- Update road intersection templates
- Keep every road line candidate
- Use the centroid of the intersections of the line candidates to
update the template
31
1 2 4 3 Blob image intersected with thinned lines 1 2 4 3 Traced line functions
SLIDE 33 Extract Accurate Road Intersection Templates
- Update road intersection templates
- Keep every road line candidate
- Use the centroid of the intersections of the line candidates to
update the template
32
1 2 4 3 Blob image intersected with thinned lines Traced line functions
SLIDE 34 Extract Accurate Road Intersection Templates
- Update road intersection templates
- Keep every road line candidate
- Use the centroid of the intersections of the line candidates to
update the template
33
1 2 4 3 Blob image intersected with thinned lines Traced line functions
SLIDE 35 Extract Accurate Road Intersection Templates
- Update road intersection templates
- Keep every road line candidate
- Use the centroid of the intersections of the line candidates to
update the template
34
1 2 4 3 Blob image intersected with thinned lines Traced line functions
SLIDE 36 Extract Accurate Road Intersection Templates
- Update road intersection templates
- Keep every road line candidate
- Use the centroid of the intersections of the line candidates to
update the template
35
1 2 4 3 Blob image intersected with thinned lines Centroid Traced line functions
SLIDE 37 Extract Accurate Road Intersection Templates
- Update road intersection templates
- Remove outliers and use the centroid of remaining intersections
36
1 2 3 4 Original map Blob image intersected with thinned line
SLIDE 38 Extract Accurate Road Intersection Templates
- Update road intersection templates
- Remove outliers and use the centroid of remaining intersections
37
1 2 3 4 Original map Blob image intersected with thinned line Traced line functions 1 2 3 4 Intersection (approximate location)
SLIDE 39 Extract Accurate Road Intersection Templates
- Update road intersection templates
- Remove outliers and use the centroid of remaining intersections
38
1 2 3 4 Original map Blob image intersected with thinned line Traced line functions 1 2 3 4 Intersection (approximate location)
SLIDE 40 Extract Accurate Road Intersection Templates
- Update road intersection templates
- Remove outliers and use the centroid of remaining intersections
39
1 2 3 4 Original map Blob image intersected with thinned line Traced line functions 1 2 3 4 Intersection (approximate location) Outlier
SLIDE 41 Extract Accurate Road Intersection Templates
- Update road intersection templates
- Remove outliers and use the centroid of remaining intersections
40
1 2 3 4 Original map Blob image intersected with thinned line Traced line functions 1 2 3 4 Outlier Centroid
SLIDE 42 Extract Accurate Road Intersection Templates
41
Previous Results This Paper
SLIDE 43 Experiments – Ground truth
- We evaluate 10 raster maps from five different sources
- Manually verify each extracted road intersection templates
with the ground truth
42
The ground truth
SLIDE 44 Experiments - Metrics
- Positional offset:
- The average number of pixels between the extracted
road intersection templates and the actual road intersections in the raster maps
43
SLIDE 45 Experiments - Metrics
- Orientation offset:
- The average number in degrees between the extract
road orientations and the actual road orientations.
44
SLIDE 46 Experiments - Metrics
- The connectivity offset:
- The total number of missed road lines.
45
Connectivity offset is one
SLIDE 47 Experimental Results
- Extracted139 road intersection templates with 438
lines from 10 test maps
- The average positional offset:
- 0.4 pixels
- The average orientation offset:
- 0.24 degrees
- Extracted road intersection templates are very close to the
ground truth
- The connectivity offset:
- We missed 13 lines from a total of 451 lines – 97% of the lines are
extracted
- Lines that do not have accurate orientations were discarded
46
SLIDE 48 Positional Offset Compared to Previous Work
47
3.66 2.44 4.52 3.14 2.7 0.12 0.52 0.37 0.69 0.35 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 Previous Work This Paper Pixel
SLIDE 49 Orientation Offset Compared to Previous Work
48
11.58 8.64 19.26 10.19 6.48 0.37 0.29 0.55 0.24 5 10 15 20 25 Previous Work This Paper Degree
SLIDE 50 Related Work
- Localized Template Matching to improve the
positional offset (Chiang et al. 08)
- The templates used for matching are not accurate
- Cluster corner points to extract road intersections
(Habib and Uebbing 99)
- Cannot extract accurate intersection templates
- Geometrical analyses to extract lines (Cao et al. 02 and
Li et al. 00)
- Do not extract intersection templates
- Color segmentation to extract lines (Khotanzad and Zink
03; Chen et al. 06)
- Do not extract intersection templates
49
SLIDE 51 Discussion
- Our technique automatically extracts accurate
road intersection templates from raster maps.
- Average positional offset: 0.4 pixels
- Average orientation offset: 0.24 degrees
- Accurate road intersection templates help to:
- Reduce searching space for map conflation application
- Use the intersection templates as seed points to extract
road from imagery
50
SLIDE 52 Future Work
- Include manual training to extract more
information from raster maps
- Labels, landmarks
- Include manual training to process more complex
maps
- A metro map with different types of lines
- Identify the training process that minimizes human
intervention
- Reuse the training results on similar maps
51
SLIDE 54 Introduction
- In our previous work [Chen et al. 2008], we extract the road
intersection templates to integrate raster maps with imagery
- Road intersection template:
- Road intersection position, connectivity, and road orientation
53
Satellite Imagery Vector data Raster map Extract Intersections Extract Intersections Extract Intersections
SLIDE 55 54
Introduction
- Label the imagery with features on the map
Tehran map from Google Image Search Satellite imagery of Tehran, Iran from Google Maps Align the map with the satellite imagery
SLIDE 56 Introduction
- Use the intersection templates as seed points to
extract road from imagery
55
SLIDE 57 Experiments - Metrics
- Positional offset:
- The average number of pixels between the extracted
road intersection templates and the actual road intersections in the raster maps
- Orientation offset:
- The average number in degrees between the extract
road orientations and the actual road orientations.
- The connectivity offset:
- The total number of missed road lines.
56
