Point Cloud Registration Based on CPD Algorithm

Abstract

A fast point cloud registration algorithm is proposed for the problem that the traditional CPD (Coherent Point Drift) algorithm is time consuming and has poor the registration efficiency. Firstly, the voxel grid method is carried out on the three-dimensional bounding box of the cloud space, and the point in the whole voxel are expressed by the voxel centers, and the down-sampling operation of the cloud is completed to reduce the amount of the calculated data. Then, the Gaussian mixture model is established for the obtained point cloud to compute the values of negative log-likelihood functions. Finally, we use the EM algorithm to iterate to solve the closed parameters by minimizing negative logarithmic likelihood function. The rotation matrix and translation vector are obtained to match two points clouds. The experimental results show that the proposed method can greatly improve the registration speed while maintaining the original registration accuracy.