Point Clouds Registration In 3D Model Reconstruction Pipeline

The three-dimensional reconstruction of real objects is an important topic and has a bright future in the field of computer vision.Generally speaking,in the process of three-dimensional model reconstruction,due to the linear propagation of light and other factors.In order to obtain a complete three-dimensional model of the object surface,it is necessary to obtain the threedimensional data of the object surface from different angles.Therefore,three-dimensional registration is an important step in the progress of three-dimensional reconstruction,which is closely related to the accuracy of three dimensional reconstruction.Up to now,the most widely used algorithm in registration is the Iterative Closest Point(ICP)algorithm,which is gradually approaching the best result by means of iteration.However,the convergence speed of ICP registration algorithm is relatively slow,and it could not achieve global optimization.Therefore,the selection of appropriate coarse registration algorithm can make the data sets obtain a relatively better initial positon,and provide a better initial position for accurate registration.The thesis mainly solves two problems in practical experiments.Firstly,we are concerned with the registration of two 3D data sets with large-scale stretches and outliers.By incorporating a weight function into the minimized error function,we got iterative reweighted least squares.In terms of calculation,this method is equivalent to weighting the point-pairs,and the weights are obtained with an M-estimation criterion.To verify the effectiveness of the proposed algorithm,we conducted comparative experiments among three algorithms: the Improved-ICP algorithm,the Scale-ICP algorithm and the standard ICP algorithm.Experiment results show that the proposed algorithm has high accuracy and strong robustness to scale and abnormal points.Secondly,in order to improve the efficiency of registration,based on the ICP algorithm,a combination of KD-tree and extrapolation is used.The advantage of the improved registration algorithm becomes more obvious as the amount of point cloud is getting greater.Experimental results show that the running time of ICP algorithm is much longer(more than 26 times)than that of the improved one,and the proposed algorithm has the advantages of fast speed and high accuracy.The experiment is based on the coarse registration algorithm of feature extraction.The initial experiment data are processed to get better initial experimental data.The experimental data after coarse registration are accurately registration and optimized.