Back to search

Democritus University of Thrace (DUTH)

Visual place recognition for robust autonomous localization in robotics applications

Abstract

dc:description

Robots have increasingly become integrated into everyday life. Autonomous vacuum cleaners are common in households, automated warehouse systems efficiently manage logistics, and self-driving vehicles and aerial drones are undergoing large-scale trials in public spaces. As these systems move beyond controlled environments, their ability to operate reliably and autonomously becomes critical. This shift places growing demands on perception, decision-making, and navigation, particularly in complex and dynamic settings. As a result, achieving robust localization remains a central challenge in autonomous robotics. A core requirement for such localization is the ability to estimate a robot's position relative to its environment using onboard sensors. While global navigation satellite systems (GNSS), such as GPS, offer reliable localization in many outdoor scenarios, they are often unreliable or unavailable in dense urban areas, forests, indoor spaces, or under heavy canopy. Alternative solutions, such as beacon or infrastructure-based localization, come with their own limitations. Consequently, robust positioning requires sensor fusion from onboard modalities such as cameras, lidars, and inertial measurement units to operate across diverse and degraded conditions. A key enabler of long-term autonomy in such scenarios is the ability to recognize previously visited locations based solely on visual input, a task known as Visual Place Recognition (VPR). VPR supports both map construction and localization by associating current observations with past ones, even when separated by significant spatial, temporal, or environmental variation. One important application of VPR is within simultaneous localization and mapping (SLAM), where it is used to perform loop closure detection (LCD). SLAM systems construct maps incrementally while estimating the robot's pose, but accumulated drift over time degrades accuracy. LCD identifies when the robot revisits a known place, allowing the SLAM backend to correct its trajectory and maintain a consistent map. In this context, VPR serves as a mechanism for spatial correction during exploration. Beyond SLAM, VPR is also a critical component within localization pipelines, where the goal is to estimate the robot's position within a pre-mapped environment. Unlike LCD, which typically operates under relatively stable conditions within a single trajectory, VPR for localization must handle significant appearance variation due to lighting, weather, seasons, and structural changes. This requires solutions that are both robust and scalable, capable of generalizing across time, space, and viewpoint, as well as operating efficiently on resource-constrained platforms. Visual localization systems must contend with a variety of real-world challenges that significantly impact performance. Environmental dynamics such as lighting, weather, and seasonal variation can dramatically alter the visual appearance of places, requiring representations that are robust to such changes. Viewpoint variation among traversals, can lead to drastic shifts in perspective that make direct matching difficult. Additionally, perceptual aliasing, where visually similar but spatially distinct locations are mistaken for one another, poses a major source of false positives, especially in structured or repetitive settings like corridors or forests. As the size of the environment increases, scalability becomes a critical issue, since methods must retrieve matches efficiently from large databases without sacrificing accuracy. At the same time, computational efficiency is essential for real-time deployment on mobile and aerial platforms with limited processing power. Finally, long-term deployments must cope with structural changes in the environment over time, such as construction, vegetation growth, or occlusions, which further challenge the reliability of purely appearance-based methods. This dissertation advances visual place recognition (VPR) by addressing robustness, generalization, and efficiency across both ground and aerial platforms. It does so through four complementary contributions, demonstrated in contexts ranging from short-term, high-precision LCD to large-scale, long-term aerial localization. We first address LCD in SLAM, where false positives can severely compromise the generated map. The proposed pipeline incrementally constructs a vocabulary from tracked binary features, retaining only long-lived tracks to ensure stability and compactness. An online co-visibility graph groups images by shared long-lived features. A community detection algorithm (CDA) then isolates locally consistent trajectory segments, improving match relevance and limiting aliasing. A temporal peak-trace filter validates loop hypotheses by analyzing the evolution of similarity scores across successive frames. Together, these components achieve perfect precision with high recall while reducing the vocabulary footprint by an order of magnitude compared to other incremental bag-of-words (BoW) methods, enabling efficient deployment on resource-constrained platforms. We next address sequence-based representation methods for VPR in long-term localization. Rather than modifying or retraining image encoder, we propose a descriptor-agnostic framework that remaps similarity scores between image sequences into a normalized distance-space domain. This representation filters out appearance-related noise while highlighting structural consistency across traversals. By exploiting temporal continuity, the method improves robustness under severe environmental changes and achieves competitive performance with state-of-the-art sequence-based representation methods, without dependence on any particular descriptor. We then address aerial VPR, a domain where models trained on ground imagery generalize poorly due to severe domain shift, rotational ambiguity, and sparse structural cues. To that end, we introduce LASED, a large-scale dataset of approximately one million aerial images collected over a decade, designed to capture geographic diversity and long-term variation. Building on this foundation, we mitigate rotational ambiguity by employing steerable convolutional networks that encode rotation equivariance directly into the architecture. This combination of structured data and specialized models substantially improves robustness to orientation changes and seasonal variation, advancing the state-of-the-art in aerial VPR. Finally, we demonstrate the practical impact of aerial VPR in a real-world deployment for unmanned aerial vehicle (UAV) localization in an open-pit mining environment. This setting involves large viewpoint changes, seasonal variation, and structural modifications, reflecting the challenges of dynamic, GPS-degraded industrial sites. Models trained on LASED consistently outperform those trained on smaller aerial datasets, highlighting the importance of scale and temporal diversity for generalization. Rotation-equivariant architectures further improve robustness to orientation changes. These results validate that curated dataset design and architectural choices enable reliable aerial localization beyond controlled benchmarks. Across these scenarios, VPR serves as the critical link between current perception and prior experience, supporting both map building through LCD and robot localization within existing maps. The contributions of this dissertation strengthen VPR in both roles: ensuring precision in SLAM loop closures, enhancing robustness for long-term localization through sequence-based representations, enabling generalization in aerial domains with a large-scale dataset, and validating these advances in a real-world UAV deployment. Taken together, the work advances VPR as a reliable foundation for autonomous navigation across ground and aerial platforms, bridging the gap between controlled benchmarks and the complexities of real-world operation.

Degree

thesis:*
Grantor dc:publisher
Democritus University of Thrace (DUTH)
Year dc:date
2025

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Papapetros, Ioannis-Tsampikos
  • Παπαπέτρος, Ιωάννης-Τσαμπίκος

Subjects

dc:subject × 26

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/60905

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Papapetros, Ioannis-Tsampikos; Παπαπέτρος, Ιωάννης-Τσαμπίκος. Visual place recognition for robust autonomous localization in robotics applications. Democritus University of Thrace (DUTH), 2025. http://hdl.handle.net/10442/hedi/60905