A novel integrated scheme for extracting superquadric-based geons from 3D data

We develop a processing scheme of volumetric primitives recognition for 3D recognition system based on Recognition by Components (RBC) theory, which integrates the advantages of deformable superquadric models reconstruction and support vector machines (SVMs) multiclass classification for extracting superquadric-based geon description directly from the shape information of 3D data. First, superquadric fitting of 3D data with real-coded genetic algorithm (RCGA) is performed for reconstructing the superquadric description of volumetric primitives; second, a novel sophisticated feature set is derived from the superquadric parameters for SVM-based classification; then, directed acyclic graph support vector machines (DAGSVM) trained by Sequential Minimal Optimization (SMO) algorithm is utilized for recognizing geon classes. Experimental results obtained show that our method is efficient and precise for superquadric-based geons extraction from real shape data in 3D object recognition.