Abstract: Aimed at addressing the shortcomings of the traditional point cloud registration normal
distribution transform (3D-NDT) and iterative closure points (ICP) algorithms, such as poor registration
effect, long registration time and serious errors, a precise and relatively efficient point cloud matching
algorithm was proposed. First, the 3D-Harris algorithm was used to identify the key points of each point
cloud, and the key points were adopted to establish a local reference frame for the basic points and
calculate the fast point feature histograms (fpfh) descriptor. Then, the corresponding estimation
algorithm of the least median of squares (LMeds) minimum median method was utilized to eliminate
the inaccurate point correspondence and obtain the feature point pairs with corresponding 3D feature
relationships. The transformation matrix required for coarse registration was calculated to complete the
preliminary registration. Subsequently, according to the 3D-NDT algorithm, the point cloud data space
was voxelized, and the probability distribution function was employed to complete the final point cloud
for accurate registration. Finally, we used this method to match three groups of point cloud files, which
were downloaded from the network with less noise and large-scale overlapped with more noise
collected by Kinect V2.0 to the same spatial reference frame, and compared the classical 3D-NDT, ICP
and other algorithms through accuracy analysis. The experimental results show that the proposed
algorithm can achieve the high accuracy registration of point cloud data in indoor scenes with low
iteration times and is less affected by noise. However, how to reduce the complexity of the algorithm
appropriately and shorten the registration time needs further research.

Key words: 3D normal distributions transform, 3D-Harris key points, fast point feature histograms;
least median of squares, point cloud registration