Research On Bayesian Rigid Point Set Registration

In the computer vision field, we usually have to measure the surface of the object many times in order to get its3D model. The data we each time measured may become rotating malposition or translational displacement or scaling of dislocation caused by the limitation of environment and instrument factors, so we must integrate those data, namely, the point set registration, which is a core problem for it plays the role of foundation of all post-processing in the computer vision field.The target of point set registration is matching the point set of two musters and estimating the matching parameters, including rotation parameters, translation parameters and scaling parameters. The algorithm of Coherent Point Drift, which is recently proposed by the PAMI(IEEE Transactions on Pattern Analysis and Machine Intelligence), is much outstanding than many other published registration methods for its accuracy. In this thesis, the author would improve the CPD algorithm in the Bayesian framework, and initially put forward logarithmic double exponential distribution. A large number of simulation experiments show that our proposed method is not only faster than the speed of the CPD algorithm, but also keeps the accuracy. Especially in the large translational displacement of the case, our method is significantly faster than CPD algorithm.The author’s main researches are as follows:(1) Research and analysis of a variety of point set registration methods and compare their advantages and disadvantages, then provide a clear sequence of the development of point set registration.(2) Study the properties of logarithmic double exponential distribution, and use it into the point set registration under the framework of Bayes.(3) Proposed using Bayesian methods to solve the problem of point set registration, and reduce the number of iterations while still keep the accuracy of the CPD algorithm.