Abstract
Road network extraction from remotely sensed imagery enables many important and diverse applications such as vehicle tracking, drone navigation, and intelligent transportation studies. There are, however, a number of challenges to road detection from an image. Road pavement material, width, direction, and topology vary across a scene. Complete or partial occlusions caused by nearby buildings, trees, and the shadows cast by them, make maintaining road connectivity difficult. The problems posed by occlusions are exacerbated with the increasing use of oblique imagery from aerial and satellite platforms. Further, common objects such as rooftops and parking lots are made of materials similar or identical to road pavements. This problem of common materials is a classic case of a single land cover material existing for different land use scenarios.
This work addresses these problems in road extraction from geo-referenced imagery by leveraging the OpenStreetMap digital road map to guide image-based road extraction. The crowd-sourced cartography has the advantages of worldwide coverage that is constantly updated. The derived road vectors follow only roads and so can serve to guide image-based road extraction with minimal confusion from occlusions and changes in road material. On the other hand, the vector road map has no information on road widths and misalignments between the vector map and the geo-referenced image are small but nonsystematic. Properly correcting misalignment between two geospatial datasets, also known as map conflation, is an essential step.
A generic framework requiring minimal human intervention is described for multispectral image road extraction and automatic road map conflation. The approach relies on the road feature generation of a binary mask and a corresponding curvilinear image. A method for generating the binary road mask from the image by applying a spectral measure is presented. The spectral measure, called anisotropy-tunable distance (ATD), differs from conventional measures and is created to account for both changes of spectral direction and spectral magnitude in a unified fashion. The ATD measure is particularly suitable for differentiating urban targets such as roads and building rooftops. The curvilinear image provides estimates of the width and orientation of potential road segments. Road vectors derived from OpenStreetMap are then conflated to image road features by applying junction matching and intermediate point matching, followed by refinement with mean-shift clustering and morphological processing to produce a road mask with piecewise width estimates.
The proposed approach is tested on a set of challenging, large, and diverse image data sets and the performance accuracy is assessed. The method is effective for road detection and width estimation of roads, even in challenging scenarios when extensive occlusion occurs.
Library of Congress Subject Headings
Remote-sensing images--Data processing; Roads--Remote sensing; Optical pattern recognition; Multispectral imaging
Publication Date
3-30-2015
Document Type
Dissertation
Student Type
Graduate
Degree Name
Imaging Science (Ph.D.)
Department, Program, or Center
Chester F. Carlson Center for Imaging Science (COS)
Advisor
Anthony Vodacek
Advisor/Committee Member
Nathan D. Cahill
Advisor/Committee Member
John P. Kerekes
Recommended Citation
Chen, Bin, "Multispectral Image Road Extraction Based Upon Automated Map Conflation" (2015). Thesis. Rochester Institute of Technology. Accessed from
https://repository.rit.edu/theses/8590
Campus
RIT – Main Campus
Plan Codes
IMGS-PHD
Comments
Physical copy available from RIT's Wallace Library at G70.4 .C44 2015