Abstract: Point cloud alignment is one of the important steps in point cloud data processingand coarse alignment is the hard part.In recent yearsgreat progress has been made in point cloud alignment based on deep learning.In particularthe method of 3DMatch can achieve a better alignment effect under the conditions of noiselow resolution and missing data.Howeverthis method uses random sampling to generate alignment points.When the number of sampling points is smallthe matching rate will be low and the alignment effect is not good.ThereforeISS feature point detection is used instead of random samplingand then 3DMatch is used to generate descriptors for feature points.Finallydata alignment is achieved through matching feature descriptors.Since ISS feature point detection has good repeatability and 3DMatch can provide descriptors with high degree of discriminationthis method greatly improves the robustness and accuracy of matching.Eexperiment shows thatcompared with random samplingthe alignment effect and robustness of feature point sampling are better when the initial point cloud has no noiseweak noise and strong noise.Moreoverwhen the coarse alignment effect is similarthe number of feature points only accounts for 10% of the number of random pointswhich greatly improves the alignment efficiency.

Key words: 3DMatch, Coarse alignment, Deep learning, Feature point detection, Point cloud alignment