Localization of a multi-articulated 3D object from a mobile multisensor system

Abstract

This paper details a model-based method for identifying and locating 3D polyhedral multi-articulated objects from a multisensory perceptual system; this system is composed of a 3D laser range finder and a CCD camera, and is mounted on a mobile robot. The method implies first to identify and locate at least one body of the object from the initial robot position, then to search the configuration of the other bodies from other positions if needed.

The initial body recognition algorithm is based on the region-edge-vertex models and the aspect graphs of the bodies, which compose the object; it relies first on a generation of hypotheses performed only with 3D data extracted from the depth map, then on a verification of each pertinent hypothesis, using only 2D data extracted from the reflectance image. Localization of the other bodies can imply to move the perceptual system around the object in order to perceive the joints and to infer relative positions between connected bodies.

In the paper, we describe algorithms developped in order to extract pertinent 3D features from such a perceptual system, to identify and localize a body, and to estimate joints configuration; experimental results performed on our mobile robot Hilare, allow to validate this approach.

This work was carried out under CNES fundings