Stratégies de perception par vision active pour la reconstruction et l'exploration de scènes statiques

Abstract

This thesis deals with the 3D structure estimation and exploration of static scenes using active vision. Our method is based on a structure from controlled motion approach which consists in constraining the camera motion in order to obtain a precise and robust estimation of the 3D structure of a geometrical primitive such as segments and cylinders. To this continuous aspect of the reconstruction process, it is necessary to define perception strategies to ensure the complete reconstruction and exploration of the scene, assumed to be composed of segments, polyhedrons and cylinders. This scene reconstruction level is essentially events/datas driven. Since this approach involves to gaze on the considered primitive, we present a method for connecting up many estimations in order to recover the complete spatial structure of scenes composed of cylinders and segments. We have developed perceptual strategies able to perform a succession of robust estimations without any assumption on the number and on the localization of the different objects. The first step of our exploration process, which includes the 3D reconstruction of a primitive, allows an incremental reconstruction of all the objects observed by the camera. We call this step local exploration due to the fact that it uses only local available informations. It is based on a prediction/verification scheme managed using Bayesian networks. This approach allows to obtain a high level representation of the considered objects, while taking into account local occlusion problems. However, when all the observed primitives have been reconstructed, a different strategy must be developed in order to ensure the completeness of the reconstruction. A global exploration process, centered on current visual features and on the structure of the previously studied primitives, is presented. This leads to a gaze planning strategy that mainly uses a representation of known and unknown areas as a basis for selecting viewpoints. Finally, the proposed algorithms have been specified and implemented using the synchronous language \signal. It allows to consider in an unified framework the various aspects of the application: from data-flow task specification (\signal) to multi-tasking and hierarchical task preemption (\signalgti). Experiments have been carried out on the Irisa robotic cell and have proved the validity of our approach.