Research On Key Technologies Of Scattered Point Cloud Registration In Terrestrial LiDAR

In oder to completely reconstruct the surface shape of an object, all point clouds acquired from different independent coordinate systems must be transformed into the same one. This procedure is usually referred as Point Cloud Registration(PCR) that is a very important part in point cloud processing. The subsequent data processing will be directly affected by its efficiency, accuracy and reliability.PCR often use the Seven-parameter Similarity Transformation Model to express the process of coordinate transformation, and then use the least-square principles to establish the coordinate conversion parameters solving model. The methods for solving coordinate transformation parameters can be broadly divided into two categories, iterative and non-iterative solutions. The iterative solutions can be used to calculate the covariance matrix for analyzing the correlation between transformation parameters, but its deficiency is susceptible to the initial values. The non-iterative solutions are very robust, but it can not be used to analyze the the correlation between conversion parameters. Therefore, a method called Solving Conversion Parameters Based on Non-iterative and Iterative Solutions(referred to as Combined Solution) has been proposed. The Combined Solution use the non-iterative solution to obtain the initial transformation parameters, and then use the iterative method to estimation the optimal transformation parameters.The Iterative Closest Point(ICP) algorithm is widely used in accurate pair-wise registration. Its key step is K Nearest Neighbors(KNN) search whose calculating time is the main factor affecting the efficiency of ICP. To improve the efficiency of KNN search, an algorithm called Orthogonal-axes Grid Partition(OGP) has been proposed based on Tree Structure and Grid Partition algorithms. The point set is projected onto a2D plane along its minimum variance direction, and then its all points are storaged based on the grid block on the2D plane, and finally the KNN of given points are searched in accordance with Grid Partition algorithm.In multi-view registration, the Constraint-based Registration algorithm can be used to allocate the accumulated error of Sequential Registration to improve registration accuracy and reliability. However, the existing methods, which did not consider the scale parameter or separately calculate the misclosure of translation parameters and rotation parameters, and then allocate the misclosure, has lower registration accuracy. An algorithm called Progressive Registration has been proposed. Its main processes are using single-site point cloud as the registration unit and building closed loop conditions of the transformation parameters obtained by ICP, and then calculating the optimum conversion parameters by conditional adjustment.Compared with the existing methods, the effectiveness and superiority of the proposed new methods are verified by experiments of measured scan data and conclusions are as follows:(1) for the Combined Solution, the non-iterative solutions (including four methods, Singular Value Decomposition, Orthonormal Matrices, Unit Quaternions, Dual Number Quaternions) and the iterative solutions (including three methods, Euler-angle Forms, Matrix Forms, Quaternion Forms), the compare experiment shows that in case of large rotation angles, the conversion accuracy of the four non-iterative solutions are consistent, the three iterative solutions converge to the wrong solutions and the Combined Solution converges to the correct solutions, and its conversion accuracy is consistent with the non-iterative solutions.(2) On the ICP convergence accuracy, the res the correct solutions, and its conversion accuracy is consistent with the non-iterative solutions, ults of OGP, Tree Structure and Grid Partition are nearly consistent. On the ICP convergence efficiency, OGP is optimal, Grid Partition is followed, and both are significantly better than the Tree Structure.(3) in terms of accuracy and reliability of Progressive Registration for the point cloud data with millimeter and centimeter average sampling interval, the translation parameters can achieve0.01mm error level, the rotation parameters can achieve second error level, and the scale parameter can reach10-6error level. Moreover, the error of conversion parameters is little affected by the sampling interval, but greater impacted by the degree of overlap of adjacent view. When the overlap of point cloud become larger, the rotate parameters can reach0.1seconds error level and the scale parameter can reach10-7error level, and so the accuracy is improved by an order of magnitude.