모바일 로봇을 위한 Semantic Zone 기반 3D 맵 관리 = Semantic Zone Based 3D Map Management for Mobile Robot

다국어 초록 (Multilingual Abstract)

Mobile robots operating in large-scale indoor environments such as hospitals and logistics centers require accurate 3D spatial representation to perceive complex structures. However, 3D maps incur substantial memory consumption, making it challenging to maintain complete map data within the robot's limited computational resources. Although existing SLAM frameworks implement memory management techniques such as graph pruning, keyframe culling, and working memory (WM) constraints —these approaches typically rely on geometric distance or temporal metrics for memory retrieval decisions, often resulting in inefficient data loading patterns in spatially compartmentalized environments.
To address this fundamental limitation, this thesis proposes a semantic zone based 3D Map Management method that fundamentally shifts the memory management paradigm from geometry-centric to semantics-centric control. Rather than managing map components based on geometric proximity alone, the proposed approach partitions the environment into semantically meaningful spatial units—such as lobbies, hallways, and patient rooms—and designates these zones as the primary unit of memory management. Specifically, the method establishes a systematic mapping between map signatures/keyframes and their respective semantic zones, enabling the system to dynamically load only task-relevant zones into working memory (WM) based on the robot's current location and operational requirements, while systematically unloading inactive zones to Long-Term Memory (LTM). This semantic zone based control strategy enables the system to strictly enforce user-defined memory thresholds MemoryThr.
The proposed method was implemented by integrating a zone-aware memory retrieval and unload algorithm into the RTAB-Map framework, maintaining compatibility with the existing WM/LTM hierarchical architecture. Quantitative evaluation against standard RTAB-Map memory management reveals that the semantic zone-based approach substantially reduces unnecessary signature load/unload cycles and cumulative memory utilization. The results demonstrate that semantic zone-based management maintains stable, predictable memory usage while preserving map availability for critical navigation and localization tasks.

목차 (Table of Contents)

• Ⅰ. Introduction 1
• Ⅱ. Background 4
• 2.1 3D Mapping and SLAM System 4
• 2.1.1 ORB-SLAM (Oriented FAST and Rotated BRIEF SLAM) 4
• 2.1.2 RTAB-Map (Real-Time Appearance-Based Mapping) 6
• 2.1.3 Nvblox (GPU-Accelerated Distance Fields) 7
• Ⅲ. Related Work 8
• 3.1 3D Map과 Submap 기반 접근법 8
• 3.2 Semantic information with 3D Map 12
• Ⅳ. Semantic zone 기반 3D 맵 관리 14
• 4.1 Overview 15
• 4.2 Semantic zone 16
• 4.2.1 기하학적 분할 방식 17
• 4.2.2 의미론적 분할 방식 18
• 4.3 Semantic zone 기반 메모리 관리 19
• Ⅴ. Implementation 21
• 5.1 Isaac Sim 기반 시뮬레이션 환경 구축 21
• 5.2 RTAB-Map 기반 semantic zone 메모리 관리 구현 22
• 5.2.1 계층적 3D Mapping 24
• 5.2.2 Semantic zone 기반 메모리 관리 알고리즘 구현 28
• 5.3 테스트 시나리오 29
• Ⅵ. Evaluation 31
• 6.1 기존 RTAB-Map 메모리 관리 알고리즘의 한계 31
• 6.2 Signature 로드, 언로드 누적 횟수 비교 36
• 6.2.1 시나리오(1) 비교 37
• 6.2.2 시나리오(2) 비교 40
• 6.3 MemoryThr를 더욱 작게 하였을 때의 semantic zone 42
• Ⅶ. Discussion 44
• Ⅷ. Conclusion 46
• Reference 48