A Novel Affine Invariant Object Alignment Scheme based on Migrant Principle and Genetic Algorithm

摘要

When the relative position between an object and the viewpoint is changed, the corresponding projected images can deviate significantly. This has imposed considerable challenges in shape recognition and pose estimation as it is difficult to determine whether two images are originated from different objects, or merely projections of the same object on multiple viewpoints. Past research had revealed that for a planar rigid object, its images observed from different positions could be related with the Affine Transform. On this basis, a pair of object contours can be classified as similar if they can be aligned by applying a legitimate Affine Transformation on one of them. Although the latter can be found with blind search, the amount of computation required to determine the six Affine parameters is enormous. To address this problem, the simple genetic algorithm (SGA) has been adopted to shorten the time required to determine the Affine Transform that leads to the best alignment between image pairs. Despite the moderate success, the method is sensitive to the quality of the initial population and exhibits a generally high failure rate. In this paper, we have overcome the above problems with a novel approach based on the integration of the "Migrant Principle" and SGA. We have restructured the former so that it can be applied in the context of object alignment. Experimental results demonstrated that our proposed method is capable of deducing the Affine Transform between image pairs with higher success rate and shorter convergence time.