In order to further improve the accuracy rate of large-scale and various kinds of 3D point clouds, this paper proposes a Convolutional Neural Network (CNN) that builds a relationship of K-nearest neighbor graphs in a local area. The key is to learn the relation between points. After the sample group finished sampling, the point cloud model is constructed and the point cloud is classified through learning the profound relationship between points and characteristics of the central point. Because this method integrated from the partial to the whole features, it makes this method to be sensitive to the shape and robust. The final experiment demonstrates that the precision rate of this method reached 92.5% on the public data set ModelNet40. Compared with the existing 3D point cloud classification algorithms, this algorithm could integrate local features and global features more effectively. Therefore, it will further increase the accuracy of 3D point cloud model classification.