Skip to content

Moving/Static Map-Cleaning Transfer Matrix

Last updated: 2026-05-09

Why It Matters

Moving/static perception data does not transfer directly into map-cleaning policy. A MOS model can say a point is moving now; a map builder must decide whether a point belongs in a permanent static map, a movable-static layer, a current-world dynamic layer, or a review queue.

For airside autonomy, this distinction is central. Aircraft, tugs, belt loaders, buses, cones, barriers, chocks, dollies, and people can all be static in one pass and invalid for a persistent map in the next.

Transfer Matrix

Source familyEvidence typeTransfers well to map cleaningDoes not transfer directlyAirside action
SemanticKITTI-MOS / LiDAR-MOSPoint-wise static/moving labels on road LiDAR sequencesBaseline loaders, IoU metrics, road actor motionAircraft/GSE semantics, low-speed apron motion, multi-LiDAR rigsUse for pipeline bring-up only
MF-MOS / CV-MOSRange/residual-map learned MOS with strong SemanticKITTI scoresModern learned dynamic masks and residual-feature baselinesSensor pattern changes and movable-static policyEvaluate after LiDAR-MOS/4DMOS baselines
4D-CSMulti-scan semantic and moving-state segmentation with cluster priorsTemporal consistency and object-level coherenceMap deletion thresholds and airport classesUse as a consistency reference, not a map cleaner
HeLiMOSMOS labels across four heterogeneous LiDAR sensorsSensor-transfer stress test, solid-state scan-pattern issuesAirport actors and permanent map decisionsRun per-sensor before fused-airside MOS
MOEDense moving-event labels and MED benchmarkDense dynamic stress tests, online/offline method comparisonStatic preservation and map lifecycle policyAdd PR/RR and latency metrics before map use
M-detectorPoint-stream moving-event outputLow-latency motion cue, training-data-free prefilterClass, instance, and permanent-map validityUse as advisory dynamic evidence
KTH Dynamic Map BenchmarkClean-map output comparison and PR/RR-style evaluationMap-level preservation/rejection workflowApron hazards, FOD, and movable-static classesReuse benchmark structure with local labels
ROADWork / RoSACamera work-zone objects and closed-zone masksCone/barrier taxonomy and temporary zone continuityLiDAR map points and 3D static-map labelsUse for visual temporary-zone context
Local airside logsFull sensor suite, poses, maps, and operational contextFinal acceptance and safety-case evidenceGeneral leaderboard comparabilityRequired before production map release

Map-Layer Policy

LayerIncludesPromotion ruleRemoval rule
Permanent staticBuildings, poles, stand geometry, stable markings, fixed barriersMulti-session agreement and localization benefitRemove only after confirmed construction/change control
Movable-staticParked GSE, staged carts, cones, barriers, chocks, temporary signsPromote only with explicit operations decisionQuarantine or expire after absence across sessions
Current dynamicMoving people, vehicles, aircraft, carts, sweeping equipmentNever promote to permanent mapRemove from mapping immediately but keep for current-world tracking
Hazard/FODDebris, small obstacles, unexpected objectsKeep as alert or review layer, not static mapClear only after inspection, disappearance, or human confirmation
ArtifactSpray, dust, reflections, multipath, registration outliersNever promoteFilter by sensor artifact logic and cross-sensor checks
Unknown/reviewDisagreement between methods or sessionsHuman or fleet-review decisionTime out only under defined map-governance policy

Metric Transfer

MetricSource benchmarkMap-cleaning interpretationAirside threshold hint
Moving IoUSemanticKITTI-MOS, HeLiMOS, MOEQuality of point-wise dynamic masksGate by actor/range/speed, not only mean
Static IoUMOS benchmarksRisk of over-removing static pointsMust stay high in localization-critical zones
MED latencyM-detector, MOE-style online testsTime before moving points are blocked from map integrationReport point-out and frame-out separately
PRKTH / HeLiMOS static map buildingFraction of valid static map retainedNear-zero erosion for stand geometry and markings
RRKTH / HeLiMOS static map buildingFraction of dynamic/transient points rejectedHigh near stands, crossings, and GSE lanes
Ghost rateLocal map QARemaining transient structure per tile or standDirect operational QA metric
Localization deltaLocal SLAM/localization replayWhether cleaning improves or harms localizationReject visually clean maps that hurt localization
False-free-space rateAirside safety validationCleaning implies free space where hazard existsTreat as safety-critical

Method Selection

Deployment stageCandidate methodsSuccess signalStop condition
Format bring-upLiDAR-MOS, SemanticKITTI-MOS loadersReproduced public metric codeLabel remap or pose alignment mismatch
Sensor transfer4DMOS, HeLiMOS baselines, MF-MOS/CV-MOSPer-sensor scores remain stableOne sensor dominates failures hidden by fusion
Dense dynamicsMOE baselines, M-detector, Dynablox/DODDynamic recall under crowded motionHigh static erosion or missed slow motion
Offline map cleaningERASOR, Removert, MapCleaner, KTH benchmark toolsHigh PR/RR and better localization replayCleaner deletes localization landmarks
Airside acceptanceBest two independent cleaners plus human reviewLow ghost rate, low false-free-space, stable route localizationDisagreement near aircraft, people, FOD, or temporary closures

Validation Guidance

  1. Do not use MOS IoU as the only map-cleaning gate. Add PR, RR, ghost rate, false-free-space, and localization delta.
  2. Separate "moving now" from "movable but static now" in labels and map layers.
  3. Require cross-session evidence before deleting or promoting cones, barriers, parked GSE, and aircraft-adjacent equipment.
  4. Keep rejected points and method disagreement as auditable artifacts for map QA.
  5. Use ROADWork/RoSA-style temporary-zone perception to explain why a map tile is quarantined, not to write LiDAR map geometry.
  6. For production airside maps, run at least two independent cleaners and require reviewer signoff in high-consequence zones.

Sources

Public research notes collected from public sources.