Deep learning feature-based visual SLAM

• Main Author: Yong, Duan Kai
• Other Authors: Lam Siew Kei
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
• Online Access: https://hdl.handle.net/10356/166008
• Institution: Nanyang Technological University

Visual SLAM is often used in autonomous agents for localization and mapping. A key element of visual SLAM is the feature detector-descriptor used to extract tracking keypoints from images of the environment. Classical visual SLAM algorithms rely on hand-crafted methods, which have low computational complexity but suffer from poor accuracy. Deep-learned feature detector descriptors have the potential to improve feature accuracy and robustness and thus tracking performance. This paper aims to compare the advantages and disadvantages of classical and deeplearning feature detector-descriptors for use in mono-SLAM through a review of literature and comparing the experimental performance of ORB-SLAM2 and SuperPoint-SLAM. A broad set of experiments covering multiple conditions such as dynamic scenes, illumination changes, and fast motion have been performed to determine baseline performance as well as reveal which conditions favour each detector-descriptor type.