Retour au site du LAAS-CNRS

Abstract

This paper explores the impact that landmark parametrization has in the performance of monocular, EKF-based, 6-DOF simultaneous localization and mapping (SLAM) in the context of undelayed landmark initialization. The work proposes a unified framework for points and lines that is independent of the parametrization used. Three parametrizations for points and five for straight lines are accurately described, emphasizing their similarities and differences. Monocular undelayed initialization challenges EKF because of the combination of non-linearity with large uncertainty associated to the non-measured DOF. In this regard, the goal of a good landmark parametrization is to improve the system's linearity as much as possible, improving the filter consistency, and thus achieving robuster and more accurate localization and mapping. The paper highlights and justifies the keys for satisfactory operation: inverse-distance parameters and landmark anchoring. It also defines a generalized linearity index suited for the EKF, and uses it to compute and compare the degrees of linearity of each parametrization. Finally, all parametrizations are benchmarked employing analytical and statistical tools, with simulations and real imagery data, using the standard and the robocentric EKF-SLAM formulations.

Abstract

In the framework of assistance robotics, this PHD aims at merging two channels of information (visual and auditive) potentially available on a robot. The goal is to complete and/or confirm data that an only channel could have supplied in order to perform advanced interaction between a human and a robot. To do so, we propose a perceptual interface for multimodal interaction which goal is to interpret jointly speech and gesture, in particular for the use of spatial references. In this thesis, we first describe the speech part of this work which consists in an embedded recognition and interpretation system for continuous speech. Then comes the vision part which is composed of a visual multi-target tracker that tracks, in 3D, the head and the two hands of a human in front of the robot, and a second tracker for the head orientation. The outputs of these trackers are used to feed the gesture recognition system described later. We continue with the description of a module dedicated to the fusion of the data outputs of these information sources in a probabilistic framework. Last, we demonstrate the interest and feasibility of such a multimodal interface through some demonstrations on the LAAS-CNRS robots. All the modules described in this thesis are working in quasi-real time on these real robotic platforms.

Résumé

Dans le cadre de la robotique d'assistance, cette thèse a pour but de fusionner deux canaux d'informations (visuelles et auditives) dont peut disposer un robot afin de compléter et/ou confirmer les données qu'un seul canal aurait pu fournir, et ce, en vue d'une interaction avancée entre homme et robot. Pour ce faire, nos travaux proposent une interface perceptuelle pour l'interaction multimodale ayant vocation à interpréter conjointement parole et geste, notamment pour le traitement des références spatiales. Nous décrivons dans un premier temps la composante parole de nos travaux qui consiste en un système embarqué de reconnaissance et d'interprétation de la parole continue. Nous détaillons ensuite la partie vision composée d'un traqueur visuel multi-cibles chargé du suivi en 3D de la tête et des deux mains, ainsi que d'un second traqueur chargé du suivi de l'orientation du visage. Ces derniers alimentent un système de reconnaissance de gestes par DBNs décrit par la suite. Nous poursuivons par la description d'un module chargé de la fusion des données issues de ces sources d'informations dans un cadre probabiliste. Enfin, nous démontrons l'intérêt et la faisabilité d'une telle interface multimodale à travers un certains nombre de démonstrations sur les robots du LAAS-CNRS. L'ensemble de ces travaux est fonctionnel en quasi-temps réel sur ces plateformes robotiques réelles.