A multi-feature fusion probabilistic topic model for unsupervised 3D point cloud classification

In this paper, a multi-feature fusion probabilistic topic model, called MFF-PTM, is proposed to realize unsupervised 3D point cloud classification. Our MFF-PTM consists of three key stages: 1) a novel multi-feature descriptor is designed to characterize different 3D point clouds by the combination of statistical, morphological and histogram features; 2) a Rsphere clustering algorithm is proposed to construct 3D visual vocabulary and generate the co-occurrence matrix, which can effectively avoid the initialization problem of category; 3) PTM employs the co-occurrence matrix to predict the probability distribution of a certain point cloud belonging to different category topics. The experimental results have clearly shown that the proposed MFF-PTM model can outperform the traditional PTM models with single feature description for 3D point cloud classification.