Registration Of LiDAR Point Clouds Based On Symmetry

With the rapid construction of the "smart city", objective needs of quickly acquiring a wide range of spatial data and timely updating are increasing. The 3D point clouds of objects can be automatic, continuous, and rapid captured by using LiDAR technology, which has the advantages of high speed, high resolution, activity, antijamming ability and etc. Relevant LiDAR technology contains satellite LiDAR, airborne LiDAR, vehicle LiDAR, and terrestrial LiDAR. And the data set with "multi-view, multi-level, multi-scale" can be obtained through different LiDAR platforms. To guarantee the integrity of data, the registration of multi-source point clouds with different coordinate systems should be conducted. The registration and integration of LiDAR data have great significance in the both aspects of scientific research and social practice.Currently, there are some technical difficulties in the registration of LiDAR point cloud due to the discreteness of points, the heterogeneity of data, the large volumes of data, small common areas, complex target structure and etc. Thus, we need to break the existing problems and explore new ideas for point cloud registration. Symmetry is a universal concept in the geographic scene, and geometric symmetries occur everywhere. The symmetry preserves a certain property (e.g. geometric invariance) of an object under some rigid operations. Inspired by the properties of symmetry, this research tries to break through the key points for LiDAR point cloud registration, with the principal line of symmetry detection, symmetry feature extraction, and point cloud registration. The ultimate goal is to perform a set of methodologies for the registration of multi-source point clouds with high accuracy, high reliability and high robustness. The research contents are as follows:(1) Symmetry detection of point clouds based on space transformationThe conventional symmetry types including mirror symmetry, rotational symmetry and translational symmetry and etc. Thus, the symmetry detection for point cloud should be conducted in advance. Then different registration strategies would be employed based on the detection results. This study proposes a symmetry detection process, which including point cloud sampling, projection, space transformation, curve fitting and density clustering. Firstly, the original point clouds are sampled and projected for reducing the calculation amount. Then the projected points are converted to the polar coordinate space. According to the points in the transformation space, the least squares curve fitting method and the Mean-shift density clustering method are used to detect the rotational and mirror symmetry of point clouds, respectively. In this method, quantitative evaluation for rotational and mirror symmetry of point clouds is performed by using the two indicators of curve fitting coefficient and mirror symmetry degree.(2) The point cloud registration method based on the rotational axisThis study proposes a rotational symmetry point cloud registration method, based on the central axis, which aim to solve the problems of large volumes of data, small overlap area and inapplicability of general registration methods. Firstly, an area deviation-based method is used to slice the point cloud in the vertical direction. Then the rotation center is extracted from the single sliced layer. The rotational axis can be obtained using total least squares fitting method. Through the central axes matching, the point clouds are reach the consistency in the vertical direction. Based on the results of central matching, the global registration is conducted by using the matching degree of the overlapping point clouds. Finally, the registration result can be optimized by using an ICP-based method with central axis constraints. This method is suitable for the point cloud with rotational symmetry, which has the advantages of high robustness, high accuracy, sample and without using auxiliary data.(3) The point cloud registration method based on the symmetry planeThis study proposes a mirror symmetry point cloud registration method, based on the symmetry plane, which aim to solve the difficulties in the general registration methods, extraction of conjugate features and acquisition of auxiliary data. The registration method consists of the following steps, including extraction of symmetry plane, point cloud projection, global coarse registration, and local fine registration. The symmetry plane of point cloud can be extracted by the density clustering method in the transformation space. On the basis of the obtained symmetry plane, the global registration is conducted with the affine invariant features on the symmetry plane. Finally, the fine registration is optimized by the classical ICP method. This method turns the point cloud registration to the problem of feature matching on the symmetry plane, which has the advantages of low complexity, high efficiency and high accuracy.The experiments show that, the rotational symmetry point cloud registration method can make use of the rotational axis feature, which can effectively avoid the conventional problems of total deviation and local optimum. When faced with small overlap area, this method can achieve an ideal registration result with high reliability and high accuracy. In the registration method for mirror symmetry point cloud, the global feature of symmetry plane is used to lead the registration of point cloud. On the basis of the symmetry plane, the final registration result can achieve high reliability and geometric accuracy.