Towards Efficient Visual Place Recognition Methods in Challenging Environments by Adaptive Representations

Abstract

Visual Place Recognition (VPR) is the ability to recognize a place by providing a query image of an unknown location. The goal is to identify an image from a geotagged database of street-side imagery that depicts the same location as the query. In outdoor environments, recognizing a place is challenging due to the visual differences between query and database images. To develop a robust VPR method capable of handling environmental changes, the image representation must possess high discrimination to distinguish relevant from nonrelevant features. However, the vast number of features between the query image and the dataset image complicates the computational process. The challenge here lies in finding an efficient way to represent images. The objective of this thesis is to present VPR methods that are resilient to dynamic environmental changes while also being efficient in terms of reducing computational demands. To achieve this goal, this dissertation explores how to create image representations that adaptively focus on specific image content. To this end, four contributions are proposed. The first and second contributions concentrate on developing efficient representation methods for accurate visual place retrieval and recognition systems. We propose methods for reducing the computational cost of calculating similarity between two vectors. As our third contribution, we suggest a hybrid feature that remains robust in the face of environmental changes. Subsequently, we extract valuable features from these hybrid representations to create an efficient VPR system. As our fourth contribution, instead of compelling the algorithm to learn relevant and irrelevant image examples, we propose a method that can predict unique features by learning both relevant and non-relevant features in a data-driven manner. In conclusion, the numerous experiments and analyses conducted in this thesis yield quantitative and qualitative results that are on par with the most advanced VPR and retrieval techniques.