Research On Globally Optimal Pointcloud Registration Based On Models

With the development of aerospace and space exploration, the realization techniques of space missions such as autonomous rendezvous and docking have become important research topics in the field of modern science and technology. And the relative pose estimation technology is essential among them. Point cloud registration is one of the basic problems in computer vision, which can be widely used in three-dimensional reconstruction, pose estimation, pattern recognition and other fields. Therefore, researches on point cloud registration methods are of great significance.Due to the advantages of simple configration, low power consumption and high accuracy, stereo vision systems for pose estimation based on dense point cloud registration have gotten a plenty of attention and researches with the progress of computer vision technologies. Dense point cloud registration is applied to solve attitude parameters during the course of pose estimation. Specifically, point cloud registration based on kown models can solve the relative pose estimation problems of non-cooperative targets, which mainly registrate a data point cloud that represents only part of the target object to its complete model. However, existing point cloud registration techniques mostly have defect of low accuracy, slow speed and bad robustness. As a result, it is hard to guarantee a successful registration between the target model and the mearusred data in an arbitrary position which may also contains some noises or outliners. Besides, the efficiency of registrating dense point clouds is quite low. These deficiencies prevent these current methods from the application to calculate attitude parameters in the pose estimation systems based on stereo vision.In order to overcome aforementioned shortcomings, this paper mainly aims at studying a globally optimal model-based point cloud registration method to meet the requests of application. Firstly, the point cloud registration algorithm based on the equivalent distance field are studied to guarantee the accurancy and speed of registration. It gets a locally optimal solution by means of building a continious distance field with the help of implicit B-splines surface fitting as well as employing Levenberg-Marquardt algorithm to optimize the distance error function. Then, the strategy for minimizing the non-convex registration error function based on Branchand-Bound algorithm is studied, which can conqure the weakness that local registration approaches may be trapped into locally optimal positions. Finally, a global registration algorithm based on models is proposed to ensure the global convergence and rubustness. And it embeds the global optimal registration method based on equivalent distance field to accelerate the convergence speed.The point cloud registration was settled by combining the local registration algorithm with equivalent distance field and global registration algorithm with the help of Branch-and-Bound method. Experiments are carried out on models in classical point cloud databses. The results and analyses validates that the developed method in the paper is of relatively high accuracy, fast speed and good robustness. What’ more, it can guarantees globally optimal convergence. Three-dimensional pose estimation tests are carried out on point clouds obtained by a stereo vision measurement system. These demostrate the proposed algorithm can satisfy the application requirements of pose estimation in the stereo vision mesurement system.