Visual Place Recognition for Autonomous Robots

Autonomous robotics has been the subject of great interest within the research community over the past few decades. Its applications are wide-spread, ranging from health-care to manufacturing, goods transportation to home deliveries, site-maintenance to construction, planetary explorations to rescue operations and many others, including but not limited to agriculture, defence, commerce, leisure and extreme environments. At the core of robot autonomy lies the problem of localisation, i.e, knowing where it is and within the robotics community, this problem is termed as place recognition. Place recognition using only visual input is termed as Visual Place Recognition (VPR) and refers to the ability of an autonomous system to recall a previously visited place using only visual input, under changing viewpoint, illumination and seasonal conditions, and given computational and storage constraints. This thesis is a collection of 4 inter-linked, mutually-relevant but branching-out topics within VPR: 1) What makes a place/image worthy for VPR?, 2) How to define a state-of-the-art in VPR?, 3) Do VPR techniques designed for ground-based platforms extend to aerial platforms? and 4) Can a handcrafted VPR technique outperform deep-learning-based VPR techniques? Each of these questions is a dedicated, peer-reviewed chapter in this thesis and the author attempts to answer these questions to the best of his abilities. The worthiness of a place essentially refers to the salience and distinctiveness of the content in the image of this place. This salience is modelled as a framework, namely memorable-maps, comprising of 3 conjoint criteria: a) Human-memorability of an image, 2) Staticity and 3) Information content. Because a large number of VPR techniques have been proposed over the past 10-15 years, and due to the variation of employed VPR datasets and metrics for evaluation, the correct state-of-the-art remains ambiguous. The author levels this playing field by deploying 10 contemporary techniques on a common platform and use the most challenging VPR datasets to provide a holistic performance comparison. This platform is then extended to aerial place recognition datasets to answer the 3rd question above. Finally, the author designs a novel, handcrafted, compute-efficient and training-free VPR technique that outperforms state-of-the-art VPR techniques on 5 different VPR datasets.