Registration of deformable objects using a depth camera

Abstract

This paper describes a method for registration and tracking of deformable objects from points clouds taken from depth cameras. Our method uses a reference model of the object in order to detect rigid and deformed regions in the input cloud. It is based on the fact that deformed objects normally have areas that are not affected by the deformations. These parts are found iteratively allowing to register the object using a chain of rigid transformations. Deformed regions are detected as those that do not satisfy rigidity constrains. Results show that correspondences of points belonging to both rigid and deformed regions can be accurately established with the reference model even in cluttered scenes.