Research On Coarse Registration Algorithm Based On Feature Point Matching

In the field of computer vision,3D object registration is a very important basic research.It is the basis and premise for 3D recognition,remote sensing,robot vision and many other fields.Therefore,3D object registration has wide application prospects.The registration methods of 3D objects can be roughly divided into two categories: coarse registration and fine registration.The coarse registration can provide a fast and efficient way to complete the registration in the case of two models with unknown position.At the same time,it can provide a good initial location for the fine registration,and reduce the consumption of the iterative time.The fine registration can make less errors between the models.But in the case of unknown position,it takes more time and is easy to get in the local optimal condition.In this paper,we mainly study coarse registration based on the local features.The main core processing can be divided into three steps,the detection of feature points,the description of the feature points and the matching of the features.For the last two stages of core processing,a feature point descriptor based on neighborhood rotation volume and a method of eliminating mismatch are proposed.So the process of point cloud registration is improved,the registration between various models is completed,Finally.The following works have been done.(1)For the feature point description algorithm is sensitive to noise and high time complexity.On the basis of the local coordinate system,the neighborhood points of the feature points are divided,firstly.Then the neighborhood points in each interval are rotated with each trapezoid that is projected to the projection point on the specific plane.Finally,we describe the interval by calculating the volume after rotation,and a descriptor for the representation of the multi-interval volume values of the feature points is formed.(2)In order to solve the problem of mismatch,the cause of mismatch is analyzed.One of them is that the edge feature points is not easy to express.In this paper,a method of eliminating edge feature points is proposed.Judging whether it belongs to the edge point according to the maximum angle between two points of the neighborhood point around the point,then the feature points in a certain range of edge points are removed to avoid the influence of lacking information on the description step.(3)Another reason for mismatching: a mismatch caused by a similar feature or a scene interference.The elimination method based on the combination of KMeans algorithm and split method is proposed.Using the idea of KMeans,the data are divided and divided into partial mismatch relations;On the basis of getting more correct matching relations,the standard deviation and Euclidean distance are used as the standard to eliminate mismatch.At the same time,the original registration process is improved by combining the research of this paper.