Skip to content

Airside Map Hygiene Ground Truth Protocol

Last updated: 2026-05-24

Why It Matters

Map hygiene ground truth defines what belongs in the permanent map, what belongs only in the live operational layer, and what should trigger review or inspection. Without this protocol, dynamic removal and map-change benchmarks can score well while erasing FOD, preserving parked GSE as infrastructure, or deleting safety-critical markings under occlusion.

The protocol below is designed for airport AGVS and fleet map maintenance. It turns raw captures into auditable labels, map-patch decisions, and safety-case evidence.

This map-operations page is the canonical owner for the ground-truth vocabulary, capture slices, reviewer dispositions, QA gates, and benchmark splits used to decide what belongs in the permanent airside map. The safety-validation companion consumes the same protocol for benchmark exchange fields, ASAM OpenLABEL-style annotation packaging, and acceptance outputs. Do not change layer names, reviewer dispositions, or acceptance semantics in only one copy; update this page first, then mirror any V&V-specific evidence fields in the companion.

Downstream consumerWhat it consumes from this protocol
Aggregated-map semantic segmentationcanonical hygiene labels plus release-state aliases (permanent_static, dynamic_residual, movable_static, static_transient, fod_candidate, artifact, unknown_review) for QA gates, training export, and semantic-map release evidence
Map construction pipelinecapture plan, rejected-object layer, change manifest, and pre-publication map-cleaning labels
Map publication gateshygiene-validation gate inputs: dynamic rejection, static preservation, FOD retention, unknown/quarantine report, and reviewer state
Map hygiene operational monitoringruntime telemetry fields and alerts for ghost rate, static preservation, FOD candidates, unknown area, and semantic QA drift
Airside map-hygiene regulatory evidencesafety-case traceability for static preservation, FOD handling, map changes, retained rejected evidence, canary monitoring, and rollback

Ground Truth Layers

LayerLabelIncludeExclude
Permanent staticstatic_keepFixed infrastructure, fixed signs, terminal edges, approved markingsAircraft, GSE, temporary objects
Static change candidatestatic_changeNew/removed/shifted fixed features pending approvalOne-pass occlusion or perception artifacts
Movable-staticmovable_staticParked aircraft, staged GSE, carts, stairs, cones, barriersPermanent localization anchors
Static transientstatic_transientStationary people, temporary work-zone objects, event equipment, staged non-asset clutterApproved permanent assets or movable assets with lifecycle ownership
Current dynamicdynamicMoving vehicles, people, aircraft under tow, wildlifeBase map elements
FOD/hazardhazardDebris, tools, loose material, spills, unsafe objectsStatic-map features
ArtifactartifactRain spray, reflections, bad registration, sensor noiseOperational obstacles unless persistent/confirmed
Unknown/reviewunknown_reviewAmbiguous clusters or conflicting evidenceAutomatic promotion/deletion

Semantic-Map Release-State Aliases

The operational labels above are the annotation vocabulary. Semantic-map release and benchmark pages use a stricter release-state vocabulary so map publication, training export, and dynamic-cleaning evaluation share one interface. Use the mapping below when producing semantic_map_manifest.json, moving/static benchmark slices, or map-derived training labels.

Ground-truth protocol labelRelease-state aliasExport rule
static_keeppermanent_staticEligible for permanent geometry, localization priors, and positive semantic-map training after geometry QA
static_changeunknown_review before approval; permanent_static after approvalNever auto-promote without ops/change evidence and reviewer decision
movable_staticmovable_staticQuarantine or soft-context layer; not a permanent-map positive
static_transientstatic_transientHard exclusion from permanent map; may become a training negative or review item
dynamicdynamic_residual when residual points remain in the aggregateRemove from permanent map and score ghost leakage separately
hazardfod_candidate unless the hazard is reclassified by inspectionPreserve evidence and reviewer state even if points are excluded from the static map
artifactartifactExclude from release and count against source-map conditioning quality
unknown_reviewunknown_reviewKeep out of automatic publication and route to reviewer or active learning

Do not infer the release-state alias from semantic class alone. A cone can be a fod_candidate, movable_static, or static_transient depending on ownership, placement, and evidence. A person can be moving, stationary-transient, or a sparse unknown cluster. The alias is a reviewed state variable, not a synonym for a class ID.

Capture Plan

Capture sliceMinimum requirementPurpose
Quiet baselineLow-traffic pass with clear sight linesPermanent static reference
Busy operationNormal aircraft/GSE activityMovable-static and dynamic separation
Aircraft present/absentSame stand in both states where possiblePrevent aircraft from entering static map
GSE staged/removedSame route with equipment changesMovable-static policy labels
Day/nightRepresentative lightingDetection robustness
Wet/dry/weatherSurface condition variationMarking visibility and artifact control
Construction/closureWork-zone capture with ops recordTemporary restriction and static-change labels
FOD drill or inspection recordControlled objects or verified inspection outcomesHazard handling validation
Do-not-delete adverse holdoutLocal chocks, cones, hoses, tools, loose debris, stationary people, staged GSE, and unknown sparse clusters under adverse-airside conditionsLocked validation of retained hazard evidence before publication

Annotation Units

UnitRequired fields
Point/voxellayer label, source capture, timestamp, pose quality, reviewer ID
Object instanceclass, bounding geometry, movable flag, current state, persistence
Map elementelement ID, prior version, new geometry, change type, approval state
Tilemap version, capture set, QA status, localization regression result
Hazard ticketobject class, location, image/point evidence, response status, clearance time

Removal-Handoff Fields

The annotation package should include the fields needed to regenerate the cleaner's map-hygiene sidecar, not just the final label. This keeps dynamic residual removal, static-transient quarantine, FOD handling, and artifact rejection compatible with semantic-map training exports.

FieldApplies toRequired content
source_evidence_refevery point/voxel/object decisionraw capture IDs, source-frame timestamps, pose-quality flag, calibration package
release_state_aliasevery publication/training-export decisionone of permanent_static, dynamic_residual, movable_static, static_transient, fod_candidate, artifact, unknown_review
reason_codeevery non-permanent or changed decisione.g. human_exclusion, movable_asset_quarantine, small_ground_unknown, free_space_contradiction, sensor_or_registration_artifact, static_wrong_demotion
semantic_evidenceclass-driven decisionssemantic class, detector/VFM candidate, confidence, reviewer override
temporal_evidencedynamic, movable, and stale-map decisionspass count, K-of-N window, absence votes, first_seen, last_seen, TTL
policy_evidencezone and asset decisionszone polygon, work order, asset registry ID, inspection ticket, waiver ID
downstream_permissionmap publication and training exportpermanent reference, soft localization context, review only, training positive, training negative, ignore mask, publish blocked

Reviewer tools should show these fields together. If a reviewer can see only the final keep/remove result, the dataset cannot support a professional safety case or a reproducible map-derived training corpus.

Labeling Workflow

  1. Register all captures into the airport ENU map frame and record pose-quality flags.
  2. Build a quiet baseline candidate from high-quality passes only.
  3. Annotate obvious permanent infrastructure and approved markings as static_keep.
  4. Annotate aircraft, GSE, workers, carts, cones, and other movable classes as movable_static, static_transient, or dynamic based on lifecycle ownership and motion/state.
  5. Route debris, loose tools, foreign material, and spills to hazard even if the static map cleaner would remove them.
  6. Compare current captures against the approved prior map and label candidate additions/removals/edits as static_change.
  7. Require reviewer decision for every safety-critical deletion and every route/topology change.
  8. Run localization and route regression on the cleaned candidate map before promoting a tile.
  9. Treat do-not-delete hazards as evidence-chain objects, not permanent-map objects. If a cleaner or semantic filter removes the points from the static map, the candidate must still be represented in hazard, movable_static, artifact, or unknown_review with source evidence and reviewer state.
  10. Label adverse-airside ambiguity explicitly. Dust, de-icing mist, steam, glycol film, wet-apron multipath, and retroreflector bloom must not be collapsed into artifact unless the reviewer can distinguish artifact returns from real object evidence using source frames, cross-sensor evidence, or repeat captures.

Reviewer Decision Table

Evidence patternLabelDecision
Fixed feature visible across independent passesstatic_keepKeep/promote
Prior feature absent in one view but occluded by aircraft/GSEunknown_reviewNo deletion
Prior marking absent in clear dry multi-view evidencestatic_changeReview deletion/repaint
New barrier appears with active work-zone recordstatic_change or temporary restrictionAdd restriction; promote only if permanent
New cone line appears without ops recordmovable_staticLive restriction, expire/review
Stationary person, passenger queue, or temporary crew cluster appears in a map tilestatic_transientHard exclude from permanent map; retain source evidence for review or training negatives
Small debris visible in routehazardAlert/inspection, not map feature
Moving object trail in accumulated mapdynamicRemove from static map
Sensor ghost or registration streakartifactReject from map and benchmark positives

QA Gates

GatePass condition
Sensor calibrationCamera/LiDAR/radar extrinsics current and versioned
Pose qualityGNSS/INS/SLAM residuals within map-build threshold
Inter-session alignmentControl points and static landmarks agree within tolerance
Static preservationFixed landmarks retained after cleaning
Movable rejectionAircraft/GSE not present in permanent layer
Static-transient exclusionStationary people and temporary non-asset objects absent from permanent layer
Hazard handlingFOD/hazard labels preserved as alerts
Local holdout integrityDo-not-delete/adverse-airside holdout is locked, traceable, and separated from training/tuning
Localization regressionNo unacceptable scan-to-map residual or relocalization degradation
Route regressionNo route graph change without approval
Audit completenessEvery promoted/deleted element has evidence and reviewer state

Dataset Splits

SplitContentRule
TrainCommon stands/routes and routine GSE variationMay include synthetic edits after real-label seed
ValidationHeld-out stands/routes at same airportTune thresholds and reviewer burden
TestHeld-out terminal/airport or real operational changesFinal airside transfer evidence
Safety casesFOD, work zones, hold points, aircraft-present edgesReport separately from aggregate metrics
Do-not-delete/adverse-airside holdoutTarget-airport hazards and adverse slices, including dust, de-icing mist, steam, glycol film, wet-apron multipath, and retroreflector bloom where in ODDLocked validation-only split; never use for training or threshold tuning

Metrics

MetricReport by
Static preservation ratefeature class, tile, stand
Movable-static rejection rateaircraft, GSE, cones/barriers
Static-transient hard-exclusion recallstationary people, temporary objects, work-zone clutter
Hazard retention/alert recallFOD type, size, surface, lighting/weather
False-free-space rateroute segment, stand, safety-critical zone
Change precision/recallinsertion, deletion, geometry edit, topology edit
Reviewer burdenalerts per stand/km/shift and approval rate
Time to decisioncapture to quarantine, review, publish, rollback
Localization impactATE/RPE, scan residual, inlier ratio, relocalization success

Sources

Public research notes collected from public sources.