Title Page
Contents
ABSTRACT 9
Ⅰ. Introduction 11
1.1. Research Background 11
1.2. Contributions 14
Ⅱ. Related Works 16
2.1. Map-cleaning Methods 16
2.2. Learning-based Online MOS. 17
2.3. Heuristic-based Online MOS 18
Ⅲ. AWV-MOS-LIO 20
3.1. System Overview 20
3.2. AWV-MOS with OST 22
3.2.1. Reference and Query Frame Selection 23
3.2.2. Point-to-Window Comparison 23
3.2.3. Point Motion Belief Estimation 26
3.2.4. Object Scale Test 29
3.3. AWV-LIO 32
3.3.1. IMU Preintegration 32
3.3.2. Feature Extraction 34
3.3.3. Scan-to-Submap Matching 35
Ⅳ. Experiments 37
4.1. Experimental Setup 37
4.2. Online Moving Object Segmentation Performance Evaluation 38
4.3. Static Map Construction Performance Evaluation 40
4.4. Moving Object Segmentation Ablation Study 43
4.5. Odometry Estimation Performance Evaluation 43
4.6. Odometry Estimation Ablation Study 45
Ⅴ. Conclusion 46
References 47
Abstract (in Korean) 51
〈Table 4-1〉 Selected subset from KITTI Raw dataset for MOS evaluation 37
〈Table 4-2〉 Online moving object segmentation performance 39
〈Table 4-3〉 Online moving object segmentation runtime performance 40
〈Table 4-4〉 Static map construction performance on KITTI Raw dataset 41
〈Table 4-5〉 Static map construction runtime performance 42
〈Table 4-6〉 Ablation study of components on the Raw 08 sequence 43
〈Table 4-7〉 Odometry estimation performance on UrbanLoco dataset 44
〈Table 4-8〉 Odometry estimation ablation study on UrbanLoco dataset 45
〈Figure 1-1〉 False segmentation case due to (a) incidence angle ambiguity (b) pose ambiguity (c) point distortion due to sensor motion. 12
〈Figure 3-1〉 System architecture of AWV-MOS-LIO. 20
〈Figure 3-2〉 Schematic illustration of AWV-MOS with OST process. 22
〈Figure 4-1〉 Qualitative result of online moving object segmentation compared with LMNet [20] on the KITTI Raw 08 sequence.... 39
〈Figure 4-2〉 Qualitative result of static map generation compared with ERASOR [18] and Removert [12] on the KITTI Raw dataset.... 42
〈Figure 4-3〉 Qualitative result of odometry estimation result on RussianHill sequence. Best viewed in color. 45