Visual localization on NTU campus

• Main Author: Chen, Wei May
• Other Authors: Lin Weisi
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering
• Online Access: https://hdl.handle.net/10356/171924
• Institution: Nanyang Technological University

This final year report presents research and development of a Visual Place Recognition (VPR) system whose innovation lies in leveraging semantic information to enhance image and information retrieval accuracy. VPR is a crucial task in computer vision and robotics, with the goal of localizing the ‘user’ of the VPR system to enable machines to understand and navigate through environments accurately. Fairly accurate image retrieval has been attainable with NetVLAD image representation. However, the presence of dynamic objects in the visual scene as well as other varying perspective and lighting conditions pose challenges to the accurate image matching of physical locations. The project explores the integration of semantic labels generated from a segmentation model to improve the image retrieval model’s performance. The main novelty of this project lies in its leverage of semantic labels to assign cluster centroids with semantic labels in order to facilitate the filtering out of dynamic objects features to enable the model to concentrate its attention and learning on the relevant, useful features without being distracted by the noisy confusing features. The experiment results demonstrate a significant improvement in the image matching accuracy, hence providing support for establishing the effectiveness of this approach.