Research On 3D Point Cloud Registration Methods

Point cloud processing technology is widely used in reverse engineering, laser remote sensing, robot vision and control, virtual reality, human-computer interaction, 3D scanning and printing, and many other fields. As the key step of point cloud processing technology,point cloud registration techniques for point cloud surface reconstruction late plays an important role. Traditional point cloud registration algorithms often require artificial and registration accuracy is not high, so, fast and accurate point cloud automatic registration algorithms have always been the research hot of point cloud processing technology. In this paper, after in-depth study of the predecessors' researches, two valid point cloud registration algorithms are accomplished in both directions from point histogram features and normal distributions transform.Firstly, the point cloud registration algorithm based on point histogram features is described. The paper introduces a new three-dimensional local feature descriptor, that is computing their normal deviation between query point and points in the neighborhood, and then construct a histogram feature,on this basis,two point clouds are aligned by using sample consensus initial alignment algorithms. Thus, initialized by the former result, the iterative closest point algorithm leads to perfect registration.Secondly, the point cloud registration algorithm based on normal distributions transform is described. The normal distributions transform can be described as a method for compactly representing a surface. It was first proposed as a method for two-dimensional scan registration.The paper extends it to apply to three-dimensional point cloud registration. Because it does not need to calculate complex feature descriptors, can greatly improve the speed of point cloud registration, but also because the use of mixed probability density function reduces the impact of external points, while greatly improving the accuracy of point cloud registration.Finally, a lot of point cloud registration experiments have been done to test the effectiveness and feasibility of the two algorithms. The point cloud data from the same experimental models are tested, the registration results of the two algorithms are compared and analyzed in detail. And the influences of different sampling rates of the two registration algorithm are analyzed to sum up the experiences for the future research.