Large Data Sets Registration Based On Feature Extraction And GMM Algorithm

Data sets registration aims to search the correspondence between two data point sets, and then match them. Because the iterative closest points (ICP) algorithm has the advan-tages of fast speed and high computational efficiency, it becomes a dominative method in the field of data sets registration. However there is some shortages in the ICP method, such as poor alignment stability, bad anti-noise ability and falling into local optimum easily, in practical it is limited to be used. Thus, how to design a more efficient and robust registration algorithm becomes an important issue in the field of image processing and computer vision. Based on the Gaussian Mixed Model (GMM), we first consider a data set as the center point of the GMM, and then find the corresponding clusters in another data set by classifi-cation methods. Furthermore, expectation-maximization (EM) algorithm is applied to solve the model, which improves the robustness and anti-noise ability obviously. However, the GMM algorithm run slowly in the case of large point sets registration problem because com-puting the correspondence of all points between two sets is time-consuming. Therefore, we consider improving the speed though feature extraction. Finally, some experimental results illustrated that the speed of the proposed method are improved.