CNN-based detector-free geometric verification for visual place recognition

• Main Author: Ji, Zhongwei
• Other Authors: Wang Dan Wei
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision; Engineering::Electrical and electronic engineering::Control and instrumentation::Robotics
• Online Access: https://hdl.handle.net/10356/165138
• Institution: Nanyang Technological University

Visual place recognition (VPR), which is essential for simultaneous localization and mapping (SLAM), is a highly challenging task in robotic systems, as it must deal with unpredictable and varied changes in the appearance of places. VPR methods can be divided into two parts: global retrieval which retrieves candidate images from the dataset, and local geometric verification which performs accurate localization. This dissertation proposes a new detector-free model for local geometric verification named Rule-based Geometric Verification (RGV). In the proposed model, the local image descriptors extracted by pre-trained convolutional neural networks (CNNs) are processed to mine the salient regions, and then different images are matched according to the similarity of salient regions. RGV can be applied to re-rank the globally-retrieved images to obtain the best-matched images.