3D scanning devices generally produce a large amount of dense points. This paper presents a feature preserving point cloud simplification method to reduce redundant points while preserving original geometric features well. Firstly,K-mean clustering algorithm was employed to globally gather similar points in a spatial domain. By constructing a K-d tree structure for the point cloud,some nodes of the K-d tree were used as initial clustering centroids. Then,normal vector of point cloud and candidate feature points were estimated with principal component analysis method. Traversing every cluster,if feature points were contained in the cluster,the cluster was subdivided into a series of sub-clusters and the cluster was mapped to a Gaussian sphere. Finally,adaptive mean shift algorithm was employed to classify the data in Gaussian sphere and the clusters in Gaussian sphere were corresponded to the sub-clusters in the spatial domain. The cluster centroids present the simplification data. Several real object models were used to verify the effectiveness of the proposed method. The experiment results demonstrate that the proposed method generates sparse sampling points in flat areas and high density points in high curvature regions. As comparing with the non-uniform grid,hierarchical agglomerative,and K-means methods,the proposed method obtains the smallest simplification error and preserves original geometric features.