Driven by data : city-scale localisation at night

• Main Author: Nelson, Peter
• Other Authors: Newman, Paul
• Published: University of Oxford 2016
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.730284

This thesis is about facilitating the use of vast quantities of data for robotics applications, particularly for the tasks of mapping and localisation in the context of lifelong learning. Further, we tackle the specific problem of localising an autonomous vehicle at night, in urban environments, using vision. In general, robotics applications have some unique and unusual data requirements. The curation and management of such data are often overlooked. An emerging theme is the use of large corpora of spatiotemporally indexed sensor data which must be searched and leveraged both offline and online. Without careful thought to their organisation, however, this leads to vast quantities of data which quickly become unmanageable. The current paradigm of collecting data for specific purposes and storing them in ad-hoc ways will not scale to meet this challenge. In the first part of this thesis, we present the design and implementation of a data management framework that is capable of dealing with large datasets, and provides functionality required by many offline and online robotics applications. We systematically identify the data requirements of these applications and design a relational database that is capable of meeting their demands. This database framework is then used to facilitate the storage and retrieval of maps in our visual nighttime navigation system, a problem which has its own unique set of challenges. Despite it being dark exactly half of the time, surprisingly little work has addressed these challenges. A defining aspect of nighttime urban scenes is the presence and effect of artificial lighting. By building a model of the environment which includes a representation of the spatial location of every light source, localisation becomes possible using monocular cameras. One major challenge we face is the gross change in light appearance as a function of distance, due to flare, saturation, and bleeding. To overcome this, we model the appearance of each light as a function of vehicle location, using this to inform our data association decisions and to regularise the cost function which is used to infer vehicle pose. We demonstrate that our system is able to localise successfully at night, over 40km, in situations where a traditional point feature based system fails.