Determining the identities, positions and orientations of randomly placed objects in 3D space is of fundamental importance in industrial and navigational robotics.
In this thesis, we describe a 3D object representation and recognition method based on superquadric, which is very efficient in acquiring the aforementioned 3D information. To represent 3D object, we exploit CSG(Constructive Solid Geometry) tree whose primitives are represented by superquadrics and their deformations. We define the set operations efficiently by the implicit functions of superquadrics and inverse function of deformations. Complex shapes are generated by deformation such as tapering and twisting.
In recognizing 3D object, we assume the visual input to the system is range data and develop a scheme using superquadrics. Because of the redundancy of size parameters of superquadrics, object may be recognized as different superquadrics. By removing the redundancy, our scheme can recognize 3D objects reliably.