Object detection in satellite images has been a key research area in computer
vision for years, with applications in GIS database updates, urban planning,
and land use analysis. The main challenges include edge extraction, line
detection, and segmentation, as identifying buildings in complex environments
with shadows, vegetation, and other obstructions is difficult. Most building
detection algorithms are edge-based, using linear feature detection,
parallelogram structure grouping, and polygon verification with geometric and
shadow-based knowledge. Recent approaches integrate multiple algorithms
and data sources to enhance accuracy and reliability. High-resolution satellite
sensors like IKONOS have improved building detection by providing detailed
imagery of urban areas. Due to the inefficiency of manual extraction,
automated methods have been developed, with most recent techniques relying
on supervised learning. These methods use initial training data to generate
hypotheses for building locations and sizes or employ model databases for
classification and matching.
One of the major challenges in building detection is the automatic and accurate
matching of buildings in satellite and aerial imagery. This study aims to use
shadow information and HSV color representation to verify building
hypotheses generated through edge-based techniques. The approach must be
both robust and accurate for practical application.
Building detection has been a significant research topic in computer vision for
years. The main difficulties arise from the presence of objects like trees, power
lines, vehicles, and parking lots, which can occlude rooftops. Additionally,
rooftops are made of various materials with different reflectance properties,
making detection more complex.
Previous research has explored different methods to tackle this problem. One
approach used a snake-based model to extract 2D building outlines from high-
resolution IKONOS images and airborne laser scanning data. Another semi-
automated technique applied an active contour model (snakes) combined with
dynamic programming optimization, requiring a digital surface model and an
ortho-image. This method becomes more effective when a human operator
manually places seed points near the building boundaries.
Several approaches have been proposed for building detection in dense urban
areas. Peng et al. used the radiometric properties of buildings to extract their
principal contours. Another method applies wavelet transform and image
scaling to analyze both high and low-frequency components of an aerial image.
One technique utilizes multiple panchromatic images, rather than stereo pairs,
to extract flat or symmetric gable-roofed buildings and generate 3D models. It
groups lines hierarchically to form building hypotheses, verifying them by
checking for expected walls and shadows. Unlike traditional methods,
verification occurs at the final stage rather than sequentially.
