Section: New Results

Visual navigation of mobile robots

Visual navigation using mutual information

Participants : Eric Marchand, Bertrand Delabarre.

We have developed a visual navigation scheme based on the mutual information between the images acquired by an onboard camera and a visual memory to control the orientation of a vehicle during its navigation [18] .

We also proposed to extend this approach to visual servoing with vision systems that consider the unified sphere model for central cameras using a normalized version of the mutual information. This permitted to apply the technique to large fields of view with a more reliable similarity function [30] .

3D Mapping and real time navigation

Participants : Maxime Meilland, Patrick Rives.

This study was realized in collaboration with Andrew Comport from I3S in Sophia Antipolis. Our approach relies on a monocular camera on board the vehicle and the use of a database of spherical images of the scene acquired during an offline step [14] . This geo-referenced database allows us to obtain a robust drift free localization. Basically, the database is composed of spherical images augmented by depth that are positioned in a GIS (Geographic information system). This spherical robot centered representation accurately represents all necessary information for vision-based navigation and mapping [37] . During the online navigation, the vehicle pose is computed by aligning the current image acquired by the camera with the closest reference sphere extracted from the database [26] .

Indoors Slam

Participants : Cyril Joly, Patrick Rives, Pierre Martin, Eric Marchand.

We developed in Sophia Antipolis a new Slam method fusing laser scan data with the spherical images provided by an omnidirectional camera. Thanks to the trace of the laser scan projected onto the spherical view, we are able to compute a RGB-D model of the environment by using a dense visual Slam approach.

In Rennes and in collaboration with Orange Labs, we considered the development of a visual Slam algorithm. Since the targeted platforms in this this study are Android Smartphone, sequential Slam approaches have been studied.

Topological navigation

Participants : Alexandre Chapoulie, Patrick Rives.

This study is realized in collaboration with David Filliat from Ensta in Paris. Navigation algorithms are often sensitive to the robot orientation involving an impossibility to detect a place already visited from a different point of view. In order to alleviate this drawback, panoramic or omnidirectional cameras are often used. We have developed a loop closure detection algorithm based on an ego-centric spherical view that satisfies, in addition to other properties, a robot orientation independence [11] .

A topological model captures the accessibility of the different places in the environment and allows a coarse localization. From a sequence of spherical views, we have developed a context-based segmentation algorithm. We hence define a topological place as having a structure which does not change, variation leading to a place change. The structure variations are detected with an efficient change-point detection algorithm [28] .

Development of an autonomous shopping cart

Participants : Luca Marchetti, Patrick Rives.

This work is realized in collaboration with Pascal Morin from Isir in Paris. It consists in developing a shopping cart with autonomy capabilities (automatic user following, obstacle avoidance, etc), as part of the Inria Large-scale initiative action Pal, which aims at developing robotic tools for disabled persons or elderlies (see Section  8.2.7 ). Experiments have been successfully conducted both on the mobile robot Hannibal and on the wheeled walking aid ANG (Assistive Navigation Guide) developed by the EPI Coprin in Sophia Antipolis [36] .

Automous navigation of wheelchairs

Participants : Rafik Sekkal, François Pasteau, Marie Babel.

This study is aimed at designing a robotic vision-based system dedicated to assisted navigation of electrical wheelchair in an unkown environment. In particular, going through doors, taking the elevator in a secure way without risking collision because of hazardous wheelchair motions remain a relevant issue. The idea is here to provide an embedded and flexible system able to ensure the immediate compatibility of the proposed system with existing electrical wheelchairs. From the platform described in Section  5.5 , we first addressed the door detection issue for automatically initializing the tracking process that is required for localisation and navigation purposes. We then defined a low complex solution of automatic door recognition that can be decomposed into three successive steps: line extraction (LSD-based algorithm), vanishing point estimation and door recognition itself by using geometrical cues. As soon as a door is detected and tracked through model-based trackers, the idea is to take into account the position of the wheelchair joystick in order to interpret the intention of the user. First experiments have shown the validity of the proposed approach. This study is conducted in conjunction with the scope of the Inria large-scale initiative action Pal (see Section  8.2.7 ).

Obstacle avoidance

Participants : Fabien Spindler, François Chaumette.

This study was realized in collaboration with Andrea Cherubini who is now Assistant Prof. at Université de Montpellier. It is concerned with our long term researches about visual navigation from a visual memory without any accurate 3D localization [9] . In order to deal with obstacle avoidance while preserving the visibility in the visual memory, we have proposed a control scheme based on tentacles for fusing the data provided by a pan-tilt camera and a laser range sensor [16] .