Muhammad Haris, "Visual Localization and Mapping in Seasonally Changing Outdoor Environments", UAS Frankfurt, 2023.

Visual localization and mapping refer to locating an agent in a scene and creating a representation of its surroundings using a camera as the primary source of perception. Localization and mapping are the fundamental prerequisites for autonomous robots, self-driving cars, and augmented reality applications. Despite decades of research and development in this domain, even state-of-the-art vision-based approaches strug- gle to perform in challenging outdoor conditions (e.g., lighting, structural, weather, and seasonal changes). Hence, this remains an active area of research and is of paramount importance for achieving autonomy over long periods. In addition to changing conditions, run-time and hardware constraints are crucial for practical ap- plications. State-of-the-art methods often rely on 3D scene reconstruction for local- ization. However, 3D scene reconstruction is resource-intensive in terms of hardware requirements and computation time. Therefore, running it on a robot equipped with low-cost hardware is infeasible. The work presented in this thesis addresses these challenges, thus providing a robust, low-cost solution for mapping and localization in outdoor environments. This thesis approaches the problem using a bio-inspired model based on unsuper- vised Slow Feature Analysis (SFA). The model reproduces the firing characteristics of Place and Head-Directions Cells found in the rodent’s hippocampus. Recently, this model has been successfully used for outdoor localization. However, it is short-term stable w.r.t environmental changes, which limits its use over a long time. Moreover, it does not scale to large-scale environments, which limits its applicability to small settings. This work overcomes the first limitation by restructuring the long-term data to change the input data’s perceived statistics. The restructuring allows the model to learn invariance to long-term scene changes. For large-scale localization, the pro- posed approach uses landmarks in a scene and learns to localize relative to them. The work also compares and analyzes the SFA-based approach with state-of-the-art methods w.r.t to localization accuracy, run-time, and hardware requirements. Suc- cessful localization and mapping enable many downstream tasks, like goal-directed navigation. The spatial smoothness of learned representation using SFA allows the approach to perform navigation using straightforward gradient descent without ex- plicit path planning. All the experiments presented in this work were performed us- ing real-world robot recordings, which enforces the feasibility of using the proposed method for practical applications.