Multi-objects interpretation

Abstract

We describe a general-purpose method for the accurate and robust interpretation of a data set of p-dimensional points by several deformable prototypes. This method is based on the fusion of two algorithms: a generalization of the iterative closest point (GICP) to different types of deformations for registration purposes, and a fuzzy clustering algorithm (FCM). Our method always converges monotonically to the nearest focal minimum of a mean-square distance metric, and experiments show that the convergence is fast during the first few iterations. Therefore, we propose a scheme for choosing the initial solution to converge to an interesting local minimum. The method presented is very generic and can be applied: (a) to shapes or objects in a p-dimensional space, (b) to many shape patterns such as polyhedra, quadrics, splines, (c) to many possible shape deformations such as rigid displacements, similitudes, affine and homographic transforms. Consequently, our method has important applications in registration with an ideal model prior to shape inspection, i.e. to interpret 2D or 3D sensed data obtained from calibrated or uncalibrated sensors. Experimental results illustrate some capabilities of our method.