Matej Artač (2006) Panoramic representations for mobile robot localisation and navigation. PhD thesis.
Computer vision enables the mobile robots to perform visual learning of the environment and the estimation of their current location. Some methods for visual learning utilise representations that consist of a series of visual impressions, captured during the exploration of the environment. In the localisation phase they compare the momentary visual impression with those stored in the representation. Thus they estimate their current location by matching the appearance. These methods perform efficiently in an arbitrary environment. However, their location estimation is limited to the vicinity of the actual training locations. In this dissertation we propose a novel method for appearance-based localisation that overcomes this limitation. The solution we propose is a synergy of a mobile robot localisation technique with a computer graphics technique. The mobile robot acquires a set of images we organise into a spatio-temporal image volume. An image-based rendering technique called the X-slits rendering uses this representation as a source for generating novel views from virtual viewpoints. We compress the representation using the principal component analysis in order to reduce the memory requirements for the representation. The proposed localisation method acquires an image from the robot's current viewpoint. It then selects the most similar image from a set of virtual viewpoints. In this way we replace the selection of pre-stored views with the rendering of the image approximations at the views with known coordinates. In order to do this, the method does not require any complex 3D reconstruction computations. Instead, it performs on a simple geometrical assumption that the scene is planar. We demonstrate the performance of the proposed method with our experiments in artificial as well as real environments of arbitrary shape. We show that the method estimates the robot's locations well despite the fact that they are arbitrary and not a part of a training set. We thus show that the proposed method is suitable for the mobile robot localisation tasks.
Actions (login required)