Visual place recognition for improved open and uncertain navigation

Abstract

Visual place recognition localises a query place image by comparing it against a reference database of known place images, a fundamental element of robotic navigation. Recent work focuses on using deep learning to learn image descriptors for this task that are invariant to appearance changes from dynamic lighting, weather and seasonal conditions. However, these descriptors: require greater computational resources than are available on robotic hardware, have few SLAM frameworks designed to utilise them, return a relative comparison between image descriptors which is difficult to interpret, cannot be used for appearance invariance in other navigation tasks such as scene classification and are unable to identify query images from an open environment that have no true match in the reference database. This thesis addresses these challenges with three contributions. The first is a lightweight visual place recognition descriptor combined with a probabilistic filter to address a subset of the visual SLAM problem in real-time. The second contribution combines visual place recognition and scene classification for appearance invariant scene classification, which is extended to recognise unknown scene classes when navigating an open environment. The final contribution uses comparisons between query and reference image descriptors to classify whether they result in a true, or false positive localisation and whether a true match for the query image exists in the reference database.