Complementarity-based Visual Place Recognition in changing environments.

Localization has long been a key research topic in computer science due to its applications in autonomous vehicles, surveillance, security, and indoor positioning systems. Visual Place Recognition (VPR), which detects previously visited locations through visual data, is a significant area within localization. Solving VPR is complex due to environmental and viewpoint variations. Despite numerous high-performing algorithms, no universal technique can address all variations with complete accuracy; each has its strengths and weaknesses for specific variations. This thesis introduces a novel element: the concept of complementarity among VPR techniques. It defines complementarity in the context of VPR algorithms and establishes the existence and degree of complementarity among existing methods. This revolutionary approach contributes significantly to developing more efficient VPR ensemble setups. It uses complementarity as a guideline to combine highly complementary techniques and avoid redundant pairings. This research introduces SwitchHit, a probabilistic, complementarity-based switching system that dynamically selects the most suitable VPR technique based on complementarity. Unlike traditional methods that run multiple techniques simultaneously, SwitchHit intelligently switches to a better technique when necessary. This feature distinguishes SwitchHit from other setups and significantly enhances VPR performance in terms of accuracy. The thesis also explores SwitchFuse, a hybrid model that combines the strengths of switching and fusion strategies for improved VPR accuracy, outperforming other similar setups, including SwitchHit and existing multi-fusion systems. The findings highlight the importance of innovative approaches, informed by complementarity analysis, to develop robust and efficient VPR systems. Additionally, it examines the utility of universal voting schemes within ensemble setups, demonstrating their potential to refine VPR accuracy further. This work lays a foundation for future research in leveraging complementarity and ensemble methods in autonomous navigation and beyond.