DC Field | Value | Language |
---|---|---|
dc.contributor.advisor | Kum, Dongsuk | - |
dc.contributor.advisor | 금동석 | - |
dc.contributor.author | Jeong, Hyeonjun | - |
dc.date.accessioned | 2023-06-26T19:32:05Z | - |
dc.date.available | 2023-06-26T19:32:05Z | - |
dc.date.issued | 2023 | - |
dc.identifier.uri | http://library.kaist.ac.kr/search/detail/view.do?bibCtrlNo=1033002&flag=dissertation | en_US |
dc.identifier.uri | http://hdl.handle.net/10203/309640 | - |
dc.description | 학위논문(석사) - 한국과학기술원 : 조천식모빌리티대학원, 2023.2,[iv, 43 p. :] | - |
dc.description.abstract | Visual place recognition is a fundamental component of various applications such as autonomous driving and mobile robotics, and many researches and developments based on deep learning have much attention by promising performance. However, it still suffers under challenging conditions from various environmental changes and the characteristics of visual sensors. Images are rich in shape, color, and texture, but information necessary for place recognition may change or be damaged due to the changes in viewpoints, appearance, illumination, and dynamic environments. Therefore, in this thesis, the additional information that complements the image and is specialized for autonomous driving is proposed to be used. Recently, map data is regarded as an important factor in autonomous driving, and it is being studied in various fields. The proposed approach performs place recognition task using map data, and the vectorized representation of map pursues the efficiency and accuracy. The data represented as vectors are no affected by information loss and performance variations, which come from rendering process. The geometric information around the vehicle can be directly used without information loss by expressing the map as a vector, and estimation can be more robust to environmental changes by focusing on a static object in the driving environment. The proposed approach trains the spatial correlation between map components represented by vectors through a graph neural network. It aims to train map components, connected by graph architecture, stably as local descriptors by hierarchical interaction. Moreover, the vectorized map-based approach is fused with existing image-based approach, and the input data work complementary to each other. The experiments demonstrate that proposed map-image fusion approach improves the accuracy of visual place recognition, and by utilizing incomplete map considering occlusion and illumination, the feasibility for autonomous driving technology can be expected. | - |
dc.language | eng | - |
dc.publisher | 한국과학기술원 | - |
dc.subject | visual place recognition▼aautonomous driving▼avectorized map▼afusion▼arepresentation learning | - |
dc.subject | 시각적 장소 인식▼a자율 주행▼a벡터화 지도▼a융합▼a표현 학습 | - |
dc.title | Vectorized map-based fusion for visual place recognition via hierarchical spatial interaction in graph | - |
dc.title.alternative | 시각적 장소 인식을 위한 계급적 공간 상호작용 기반의 벡터화 지도의 융합 | - |
dc.type | Thesis(Master) | - |
dc.identifier.CNRN | 325007 | - |
dc.description.department | 한국과학기술원 :조천식모빌리티대학원, | - |
dc.contributor.alternativeauthor | 정현준 | - |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.