Study on collision prevention technology based on LiDAR-radar data fusion

국문 초록 (Abstract)

반도체FAB 환경의OHT(Overhead Hoist Transport) 시스템은 천정 레일을 따라 FOUP를 이송하며, 협소한 주행 경로와 합류/분기 구간에서 사람·설 비·차량 등다양한장애물이등장할수있다.이때추돌은생산중단과장비 손상을유발하므로,충돌예방은사실상무관용수준의요구사항이다.현장에 서는주로패턴형2DLiDAR를이용해근접안전영역을감시한다.2DLiDAR 는고해상도거리·공간정보를제공하지만,투명체(아크릴/유리)는레이저투 과(pass-through)로 인해 미검출이 발생할 수 있고, 난반사/광택 재질에서는 반사강도변화로거리값이불안정해지는한계가있다.또한합류부·곡선부에 서는시야각제한과가림으로인해감지지연이발생할수있다.이러한한계를 보완하기 위해 차단막 설치, 반사 테이프 부착, 합류 제어 시스템 등의 추가 설비를설치하여운용하는경우가많다. 본논문이제안하는핵심기법은Radar–LiDAR트랙–투–트랙공분산교차 (T2T–CI) 융합이며, 이를 개발하기 위해 2단계 연구를 순차적으로 수행하였 다. 1단계(선행연구)에서는포인트클라우드/측정치수준의융합(PointCloud Fusion, PCF)을 통해 투명체 및 반사 환경에서 정지거리 분산을 감소시키고, 센서융합이실제운용상의한계를완화할수있음을확인하였다.다만PCF는 “거리 안정화”에는 효과적이지만, 이동 표적의 연속 추적과ID 영속성확보, 다중 표적 교차/가림 상황에서의 안정성에는 한계가 있었다. 이에 2단계(본 연구)에서는 레이더(10 Hz)와 라이다(40 Hz) 각각 독립적인 다중표적 트래커 (JPDA–IMM 기반)를구성한뒤,공통시간으로예측하여트랙을정렬하고,상 관관계를알수없는두추정치를공분산교차(CI)로보수적으로융합하였다. 트랙 매칭은χ2 기반 위치 게이트이후속도/도플러 일관성게이트를추 가하는 2단계 게이팅으로 오연합과 유령(ghost) 트랙을 줄였다. 또한 거리와 센서 품질(NumPoints, ClusterQuality), 레이더 도플러 신뢰도 및 존재확률 (existence probability)을 반영하여 위치/속도 블록을 분리해 가중치를 적응적 으로조정하는AdaptiveBlock-CI를설계하였다.더불어다중프레임히스토리 를활용한ID재사용과트랙라이프사이클(확정/삭제/병합)관리를포함하여, 가림/교차상황에서도ID영속성과유령억제를강화하였다. 정밀한Ground Truth(GT) 제작이 어려운 환경을 고려하여, 본 논문의 또 다른 핵심 기여로 GT-free 평가 체계인 SCRA(Statistical Consistency and Residual Analysis)와 요약점수FQS(FusionQualityScore)를제안한다.SCRA 는연속성(TCI),유령률(GTR)및억제율(GSR),센서간residual/NIS,물리적 TTC(TTC = r/(−˙r)), 확정 지연(TTCf), 실시간성(Runtime) 등을 통합적으 로 분석한다. 1/2/4인 보행 시나리오에 적용한 결과, 제안한 T2T–CI는 단일 센서 대비연속추적과유령억제를동시에개선하고,통계적일관성을유지 하면서도실시간실행이가능함을확인하였다.대표적으로4인시나리오에서 TCI 84.6%, GTR 5%, GSR 95%, NIS p95 ≈ 10.4, FQS 91.3을 기록하였다. 주제어:충돌방지,레이더–라이다융합,트랙–투–트랙융합,공분산교차(CI), JPDA–IMM, 물리적 TTC, GT-free 평가(SCRA/FQS)

다국어 초록 (Multilingual Abstract)

Ceiling-mounted Overhead Hoist Transport (OHT) systems in semicon ductor fabrication facilities transport FOUPs along narrow rail corridors with frequent merge/branch sections. Because a single contact incident can cause wafer loss and long downtime, collision avoidance is effectively a zero-tolerance requirement. In industrial practice, proximity sensing is largely based on pattern-type 2D LiDAR, which provides dense range and geometric informa tion. However, 2D LiDAR can miss transparent obstacles (e.g., acrylic and glass) due to beam pass-through and can produce unstable ranges on specu lar or glossy surfaces. Therefore, additional countermeasures such as physical barriers, reflective tapes, and merge-control interlocks are often installed, in creasing installation and maintenance cost. Millimetre-wave radar is robust to transparency and provides Doppler velocity, but suffers from low angu lar resolution and multipath clutter in narrow corridors. These limitations motivate complementary Radar–LiDAR fusion. This thesis presents a Radar–LiDAR track-to-track covariance intersec tion fusion framework (T2T–CI) that improves track continuity, suppresses ghost tracks, and provides physically meaningful collision-risk indicators. The study was conducted in two phases. Phase 1 (industrial pre-study) fused radar and LiDAR at the point-cloud/measurement level to stabilise stop-distance commands for transparent and reflective obstacles, validating the practical ity of sensor fusion. Phase 2 builds independent JPDA–IMM multi-target trackers for radar (10 Hz) and 2D LiDAR (40 Hz), predicts tracks to a common time, associates radar and LiDAR tracks with a two-stage gating strategy (a chi-squared position gate followed by a velocity/Doppler consis tency gate), and fuses matched tracks using covariance intersection to handle unknown cross-correlation. An adaptive block CI scheme separately weights position and velocity components using range and quality cues (LiDAR Num Points/ClusterQuality, radar Doppler reliability, and track existence proba bility). ID persistence is enhanced using multi-frame history-based ID reuse and track lifecycle management. Because dense ground truth is difficult to obtain in industrial settings, this thesis also introduces a GT-free evaluation framework—Statistical Con sistency and Residual Analysis (SCRA)—and a summary score, Fusion Qual ity Score (FQS). SCRA combines continuity (TCI), ghost-related metrics (GTR/GSR), radar–LiDAR residual and NIS consistency checks, physical time-to-collision TTC = r/(−˙r), time-to-confirm, and runtime. In a four person scenario, the proposed T2T–CI achieved TCI 84.6%, GTR 5%, and GSR95%,maintained NIS p95 ≈ 10.4, and achieved FQS 91.3, demonstrating robust tracking and interpretable safety margins under real-time constraints. Keywords : Collision Prevention, Radar–LiDAR Fusion, Track-to-Track Fu sion, Covariance Intersection, JPDA–IMM, Physical Time-To-Collision, GT Free Evaluation (SCRA/FQS)

목차 (Table of Contents)

• 1 Introduction 1
• 1.1 BackgroundandMotivation 1
• 1.2 ProblemStatement 6
• 1.3 ResearchObjectives 8
• 2 RelatedWork 12
• 2.1 Multi-SensorTrackingandFusionTrends 12
• 2.2 Track-to-TrackFusionandCovarianceIntersection 13
• 2.3 JPDAandIMMforMulti-TargetTracking 14
• 2.4 SafetyMetrics:PhysicalTime-to-CollisionandTime-to-Confirm 15
• 3 SystemConfigurationandDataPreparation 16
• 3.1 OHTEnvironmentandOperationalConstraints 16
• 3.2 SensorSpecificationsandMeasurementModels 19
• 3.3 LiDARDataPreprocessing 21
• 3.4 RadarConfiguration 22
• 3.5 TimeAlignmentandCoordinateNormalization 23
• 4 PointCloudFusionandExperiments 24
• 4.1 PCFAlgorithmFramework 24
• 4.2 Stop-DistanceCommandLogic 27
• 4.3 ImplementationDetailsandParameterSelection 29
• 4.4 PCFExperimentsandResults 29
• 4.5 SummaryandTransitiontoTrack-LevelFusion 38
• 5 T2T-CIFusion 39
• 5.1 Single-SensorTracking(RadarOnly,LiDAROnly) 39
• 5.2 Two-StageGatingforRobustAssociation 44
• 5.3 CovarianceIntersectionModel 44
• 5.4 AdaptiveBlockCIWeightingStrategy 45
• 5.5 IDPersistenceandHandoverMechanism 46
• 5.6 PipelineParametersandExecution 46
• 6 T2T-CIFusionExperimentsandResults 51
• 6.1 ScenarioDescriptionsandExperimentalSetup 51
• 6.2 BaselineResults(RadarOnly,LiDAROnly) 52
• 6.3 T2T-CI FusionResultsfor1-,2-,and4-PersonScenarios 53
• 6.4 StatisticalConsistencyandResidualAnalysis(SCRA) 58
• 6.5 SCRADetailAnalysis 60
• 6.6 FusionQualityScore(FQS) 63
• 7 Conclusion 65
• 7.1 SummaryofContributions 65
• 7.2 KeyQuantitativeImprovements 66
• 7.3 OverallConclusion 68
• Bibliography 70
• Abstract 74