An evolutionary algorithm for the registration of 3-d surface representations

The registration of 3-D surface representations is an important task for the recognition of objects and for the fusion of different views (reconstruction). Finding the transformation parameters that optimally align two non-calibrated segmented scene descriptions is a difficult, complex optimization problem. In this paper an Evolutionary Algorithm (EA) is presented which offers a solution to the registration problem. The fitness function which estimates the quality of the transformation parameters is based on single surface comparisons achieved by a neuro-fuzzy system. We demonstrate some registration experiments with synthetic and real scene descriptions, showing the robustness of the registration with respect to segmentation noise and partial visibility.