Chinese Journal of Computers   Full Text
Title	Range Image Registration Using a Hybrid Genetic Algorithm and Surface Mean Inter-Space Measure
Authors	GAO Peng-Dong1),2) PENG Xiang1),3) LI A-Meng3) LIU Ze-Yi4)
Address	1)(State Key Laboratory of Precision Measurement Technology and Instrumentation, Tianjin University, Tianjin 300072) 2)(High Performance Computing Center, Communication University of China, Beijing 100024) 3)(Key Laboratory of Optoelectronic Devices and Systems of Chinese Education Ministry, Institute of Optoelectronics, Shenzhen University, Shenzhen, Guangdong 518060) 4)(Department of Mathematics of Science College, Shenzhen University, Shenzhen, Guangdong 518060)
Year	2007
Issue	No.12(2189—2197)
Abstract &   Background	Abstract A novel approach is presented for precise registration of polygon meshes pair with a new error metric: Surface Mean Inter-Space Measure(SMISM). The method is based on an improved genetic algorithm. Unlike the existing distance-based measures, the SMISM takes on the mean 3-D inter-space associated with each triangle in the overlap region to guide the range image registration. In addition, a hybrid genetic algorithm is able to register range images without need for pre-alignment, which is a key limitation always afflicting the well-known iterative closest point (ICP) method. The proposed hybrid GA, combined with the strategy of simulated annealing(SA) selection, best individual migration and dynamic parametric space degeneration, offers much faster convergence and more precise registration than the traditional GA methods. A set of experiments is designed to demonstrate that the presented method is insensitive to noises and has high precision as well as the fast convergence. keywords genetic algorithm; surface mean inters-space measure； range image registration; error metric background The research of this paper is supported by the National Natural Science Foundation of China "Multi-resolution Dynamic 3-D Imaging and Modeling" (No.60275012), the Key Research Program of Natural Science for University in Guangdong Province (No.04Z010), the Natural Science Foundation of Guangdong Province (No.031804), and the Research Project of Science & Technology from Shenzhen Government (No.200341). The research work reported in this paper focuses on the precise registration of polygon meshes pair. In Computer Vision and Computer Aided Geometric Design literatures, 3D modeling from physical objects has been one of the most active research realms having many practical applications, such as advertisement, entertainment industry, construction of virtual museums, and so on. Since, a physical object can not be completely scanned with a single image, multiple scans from different views are required to supply the information needed to construct the 3D model. Therefore, accurately registering these range images corresponding to different views into a single 3D model has been a significant problem. And more attention has been paid to it by researchers all over the world since 90’s last century. However, the existing registration methods, based on the point-to-point or the point-to-plane distance, always suffer from all kinds of outliers. This paper presents a new volume-based error metric and a novel hybrid GA to finish the precise registration of two triangular meshes. Comparing to the distance-based measures, a number of experiments have illustrated that the presented method is effective and robust.