실외환경의 건물 외곽선 및 출입구 정보를 활용한 위성항법-시각-관성 융합 기반 의미론적 동시적 위치추정 및 지도작성 = GNSS-Visual-Inertial Fusion Based Semantic Simultaneous Localization and Mapping Utilizing Building Contour and Entrance Information in Outdoor Environments

다국어 초록 (Multilingual Abstract)

This paper proposes a GNSS–visual–inertial SLAM framework that integrates semantic doorway information and building contours that reflect the height structure of buildings, aiming to achieve robust outdoor localization and globally consistent mapping in urban environments for autonomous navigation. A doorway detection module is designed by combining a YOLO-based detector for doors and handles with a multilayer perceptron binary classifier that judges geometric and scale consistency, so that doorway detection remains stable under strong illumination changes, glass reflections, occlusions, and diverse façade materials. Experiments using a dataset collected on real campus streets show that the proposed classifier increases doorway detection precision to about 94% while maintaining a practically meaningful level of recall. This provides stable semantic object tracking and registration that can be used in the SLAM system.
The detected doors are temporally tracked using IoU based data association between consecutive keyframes and handle-layout analysis for double-leaf structures, and are represented as global doorway descriptors that integrate accumulated 3D pose, orientation, and geometric properties over time. In addition, true-north information obtained from GNSS is used to transform the camera poses so that each doorway is aligned in an NED coordinate frame together with its latitude and longitude. Through this process, the gap between the local SLAM coordinate system and an Earth-referenced coordinate system is bridged.
Furthermore, pseudo LiDAR is generated from stereo depth and the ground plane is estimated, after which building contours with layer-wise height structure are generated and projected onto a 2D ground plane. By using these contours as objects in the SLAM framework, the long-term structural stability of buildings is exploited. The generated contours are used to validate and reinforce loop closure through layer-wise 2D and 3D rigid registration and a height-weighted matching score. On urban pedestrian sequences of about 1 km, the proposed contour-based verification method is appended to the back-end of a DBoW3-only baseline for comparison. When this verification is added, false positive loop closures are reduced by about 60%, and loop-closure precision is improved from roughly 93.6% to 97.4%. Overall, integrating semantic features and structural features significantly improves the reliability of loop matching, and the results indicate that the proposed framework can be effectively applied to service robot scenarios that require navigation in real outdoor environments.

목차 (Table of Contents)

• 제1장 서론 1
• 제1절 연구 소개 1
• 제2절 관련 연구 5
• 제2장 이론적 배경 11
• 제1절 위성항법-시각-관성 팩터 그래프 기반 자세 추정 11
• 1. 상태벡터 정의 11
• 2. MAP 기반 비선형 최적화 13
• 3. 팩터 그래프 기반 자세 추정 14
• 제3장 안정적인 출입문 추적 및 등록 18
• 제1절 개요 18
• 제2절 안정적인 출입문 탐지 18
• 1. 출입문을 구성하는 객체 탐지 20
• 2. 본객체와 부분객체의 위치 및 스케일 관계 정의 22
• 3. 다층 퍼셉트론(MLP) 기반 이진 분류 모델 구성 25
• 4. 안정적인 출입문 탐지 26
• 제3절 출입문 추적 및 등록 28
• 1. 출입문 추적 29
• 2. 출입문 등록과 활용 31
• 제4장 건물 외곽선 추출 및 등록 38
• 제1절 개요 38
• 제2절 건물 외곽선 추출 38
• 제3절 건물 외곽선 기반의 루프 매칭 강화 42
• 제4절 루프 클로징 45
• 제5장 실험 결과 47
• 제1절 안정적인 출입문 검출 47
• 1. 실험 환경 47
• 2. 비교 방법들과 평가 지표 48
• 3. 정량적 평가 결과 50
• 4. 정성적 평가 결과 55
• 제2절 건물 외곽선 기반의 루프 매칭 강화 56
• 1. 실험 환경 57
• 2. 비교 방법들과 평가 지표 58
• 3. 정량적 평가 결과 59
• 4. 정성적 평가 결과 61
• 제6장 결론 65
• 제7장 참고문헌 67