| Résumé : |
(auteur) This thesis focusses on the task of navigating an autonomous vehicle along a previously driven path using feedback obtained from a camera system. The desired path is stored in the form of a “visual memory”, essentially a collection of images, captured when the vehicle was first driven along the path. Algorithms of this form find applications in many semi-autonomous inspection/exploration tasks, where the vehicle is initially navigated by an operator, using the visual data for other purposes. On completion of the task, the operator has provided the autonomous vehicle with all the information it needs to find its way back to the starting location, and potentially repeat the entire trip. By using reference images recorded along the initial path, the system is afforded a form of global localisation using only local sensing, by providing relative information to specific key-points within the environment.
The work presented in this thesis uses, as a base, two well established path following controllers, and extends these control algorithms into the visual domain, by deriving the required parameters of each of the controllers from information gathered in the images. One of the key focus points of this work is the use of only the bearing (yaw) information from the images. By only working with feature bearing information we essentially reduce the number of parameters by half (by ignoring elevation) without sacrificing performance on 2D-manifolds.
The first controller extends the well-known shortest distance to path control algorithm, by deriving a scaled distance to path and relative orientation from the visual memory. Using the scaled distance to path, we incorporate the unknown scale that typically plagues vision-based systems, into the controller to remove the velocity dependence of the control law. This algorithm was implemented and tested in an indoor environment with a motion capture system.
The second controller extends a model predictive control (MPC) based algorithm, derived during the 2007 DARPA Grand Challenge and initially reliant on GPS information, to make use of image data, thereby alleviating the need for GPS position information. To achieve this, a novel image-based cost function is proposed that can relate the relative distances between several images. This cost function guides the choice of control trajectories to minimise the computed cost from the reference path. The performance of the proposed cost function is examined in detail, including the effects of the number of features, average distance to feature, feature observation noise and the number of outliers.
To use this cost function, the control algorithm also needs an indication of how future actions will affect the cost, and to this end feature extrapolation becomes necessary. With limited visual information, and short baselines, this process is often not very successful, and various techniques are presented to improve the performance. These include the weighting of features based on their error prediction, and the reduction of the prediction horizon required by the controller.
This control algorithm was demonstrated in both urban and extra-urban settings over paths on the order of 400m where the performance is shown to be comparable to that of differential GPS. Finally, it is also shown how the algorithm can be simply adapted to incorporate collision avoidance behaviour during the path replay in the event that the environment has changed between recording and playback. |
Ajouter le résultat dans votre panier
Visionner les documents numériques
Affiner la recherche Interroger des sources externesCoregistration refinement of hyperspectral images and DSM: An object-based approach using spectral information in ISPRS Journal of photogrammetry and remote sensing, vol 100 (February 2015)
Algorithms for vision-based path following along previously taught paths (2015)
