Visual Place Recognition under Severe Viewpoint and Appearance Changes

Over the last decade, the eagerness of the robotic and computer vision research communities unfolded extensive advancements in long-term robotic vision. Visual localization is the constituent of this active research domain; an ability of an object to correctly localize itself while mapping the environment simultaneously, technically termed as Simultaneous Localization and Mapping (SLAM). Visual Place Recognition (VPR), a core component of SLAM is a well-known paradigm. In layman terms, at a certain place/location within an environment, a robot needs to decide whether it’s the same place experienced before? Visual Place Recognition utilizing Convolutional Neural Networks (CNNs) has made a major contribution in the last few years. However, the image retrieval-based VPR becomes more challenging when the same places experience strong viewpoint and seasonal transitions. This thesis concentrates on improving the retrieval performance of VPR system, generally targeting the place correspondence. Despite the remarkable performances of state-of-the-art deep CNNs for VPR, the significant computation- and memory-overhead limit their practical deployment for resource constrained mobile robots. This thesis investigates the utility of shallow CNNs for power-efficient VPR applications. The proposed VPR frameworks focus on novel image regions that can contribute in recognizing places under dubious environment and viewpoint variations. Employing challenging place recognition benchmark datasets, this thesis further illustrates and evaluates the robustness of shallow CNN-based regional features against viewpoint and appearance changes coupled with dynamic instances, such as pedestrians, vehicles etc. Finally, the presented computation-efficient and light-weight VPR methodologies have shown boostup in matching performance in terms of Area under Precision-Recall curves (AUC-PR curves) over state-of-the-art deep neural network based place recognition and SLAM algorithms.