| Summary: | 碩士 === 國立成功大學 === 測量工程學系碩博士班 === 90 === Model-based building extraction (MBBE) is currently one of the major research topic in the field of digital photogrammetry. The effectiveness of optimal model-image fitting algorithm is the key of MBBE. The previously proposed Least-squares Model-image Fitting (LSMIF) algorithm is an iterative solution using Newton’s method which needs good initial approximations of the unknown parameters and is easy to get trapped at a local optimal solution. Genetic Algorithms (GA) work with rich population and simultaneously climbs many peaks in parallel during the search process. It effectively avoids the possibility of getting trapped at a local minimum. This thesis, therefore, tailored a GA to be a new model-image fitting method for MBBE, which does not need good approximation.
In this thesis, parameterized CSG (Constructive Solid Geometry) primitives including box and gable-roof house are tested for demonstration. In the approach, the initial population is generated randomly in the predefined parameter domain. Consequently, this population of primitives are transformed to object coordinate system with pose and shape parameters and projected to photo coordinate system with the known exterior orientation parameters. We propose a concept of fitness function exploiting objective of minimization of discrepancy between the extracted edge pixels and projected model wire frame edge to evaluate each individual in population. After a series of re-product, crossover, and mutation operators, taking the model with highest fitness value as solution. Our experimental results show that GA can correctly and globally find the near-optimal solution by compared with manual measurements and LSMIF. In the conclusion, GA indeed has potential for model-image fitting.
|
|---|
