To date, computer vision methods have largely focused on extraction from panchromatic imagery. Despite significant technological advances, road extraction algorithms have fallen short of satisfying rigorous production requirements. To that end, the objective of this thesis is to present a new approach for automating road detection from high-resolution multispectral imagery.; This thesis considers three main research objectives: (1) development of a fully automated road extraction strategy in that interactive human supervision or input initializations are not required; (2) development of a globalized approach to road detection that is motivated by principles of self-organization; (3) meaningful exploitation of high-resolution multispectral imagery. Several new techniques are presented for fully automated road extraction from high-resolution imagery. The core algorithms implemented include (1) Anti-parallel edge Centerline Extractor (ACE), (2) Fuzzy Organization of Elongated Regions (FOrgER), and (3) Self-Organizing Road Finder (SORF). The ACE algorithm extends the idea of anti-parallel edge detection in a new approach that considers multi-layer images. The FOrgER algorithm is motivated by Gestalt grouping principles in perceptual organization. The FOrgER approach combines principles of self-organization with fuzzy inferencing to building road topology. Self-organization represents a learning paradigm that is neurobiologically motivated. Globalized analysis promotes lower sensitivity to fragmented information, and demonstrates robust capacity for handling scene clutter in high-resolution images. Finally, the SORF algorithm bridges concepts from ACE and FOrgER into a comprehensive and cooperative approach for fully automated road finding. By providing an exceptional breadth of input parameters, output metrics, modes of operation, and adaptability to various input, SORF is particularly well suited as an analytical research tool. Extraction results from the SORF algorithm are compiled from scenes over four different test areas, and five different sensors. The results show the highest extraction quality rates from anti-parallel edge analysis of spectral band layers, and the highest extraction correctness rates from anti-parallel edge analysis of spectral class layers. Spectral class layer analysis generally requires a lower computational requirement per layer. When edge analysis is rendered ineffective due to excessive scene clutter, a strictly regional analysis of spectral class layers can provide a more effective means of extraction.
展开▼
