Point cloud registration algorithm based on NDT with variable size voxel

To improve the accuracy of point cloud registration, this paper proposes a method of point cloud registration using variable size voxel based on normal distributions transform (NDT). Firstly, voxels with large size are used to segment point cloud. And then depending on the distribution-density of points segmenting, the large voxels are segmented into several voxels with small size. So it can aggregate the sparse points into a big voxel and disperse the dense points into multiple small voxels, which can eliminate large different of number of points among voxels with fixed size and avoid the defect that some sparse points can´t be used. Secondly, mixed probability density function is designed which combines a uniform distribution function with the normal distribution function to enhance robustness of registration of point cloud with noise. Experiments verifies that the proposed registration algorithm with variable size voxel can get better registration accuracy than the fixed size voxel, while the mixed probability density function has stronger anti-noise ability than the single probability density function.