Skip to content

Airside Dynamic Map Cleaning Benchmark

Last updated: 2026-05-09

Why It Matters

Airside dynamic map cleaning is a safety validation problem, not only a perception benchmark. A cleaned map can affect localization, route availability, obstacle expectations, and safety-case evidence for Autonomous Ground Vehicle Systems (AGVS). The validation set must therefore combine map-cleaning metrics with airside hazards such as FOD, ground crew, aircraft, baggage carts, tugs, temporary barriers, and apron lighting/weather.

FAA material makes the operating context explicit: FOD is a continuing airport safety concern, AGVS include FOD detection/retrieval and airside service vehicles, and AGVS testing must be coordinated and controlled before moving toward autonomy without human supervision.

Dataset/Benchmark Table

Source / benchmark inputSource URLWhat it providesHow to use in an airside map-cleaning benchmarkLimitation
FAA Foreign Object Debris Programhttps://www.faa.gov/airports/airport_safety/fodAuthoritative FOD safety definition and references to FOD management and detection equipment advisory circularsDefine hazard taxonomy, object-size/material classes, inspection expectations, and false-free-space severityNot an ML dataset; use as safety and acceptance guidance
FAA automated FOD detection system evaluationhttps://www.airporttech.tc.faa.gov/Airport-Safety-OLD/Airport-Safety-and-Surveillance-Sensors/Automated-Foreign-Object-Debris-FOD-Detection-System-EvaluationHistorical FAA evaluation context for radar/electro-optical FOD detection systemsBorrow detection-probability, false-alarm, response-time, and operational-coverage thinking for autonomy perceptionFocused on FOD systems, not dynamic map cleaning
FOD-Ahttps://github.com/FOD-UNOmaha/FOD-dataAirport runway/taxiway FOD images with bounding boxes plus light-level and weather annotationsValidate FOD visual detection and negative controls for map cleaning around small objectsCamera image dataset; no 3D map labels, localization, or LiDAR occupancy
AIT Apron Datasethttps://publications.ait.ac.at/de/datasets/apron-dataset/Airport-apron logistics image dataset with bounding boxes, object categories, and environmental meta-parametersBuild airside object taxonomy and detection robustness slices for GSE/personnel/apron objectsImage-focused; does not provide dynamic 3D map ground truth
FAA AGVS on Airportshttps://www.faa.gov/airports/new_entrants/agvs_on_airportsCurrent FAA AGVS context, applications, contacts, and published information, including Bulletin 25-02 and CertAlert 24-02Define regulatory validation envelope, test-plan evidence, and controlled-environment constraintsDoes not define technical perception metrics
FAA Emerging Entrants Bulletin 25-02https://www.faa.gov/airports/new_entrants/bulletins/25_02Guidance for testing/demonstrating AGVS, including human monitor, control capability, route/test plan, RF/aeronautical-study considerationsConvert benchmark runs into safety-case evidence with test plans, roles, mitigations, and human takeover assumptionsGuidance is operational/regulatory, not a benchmark dataset

Metrics

MetricDefinitionAirside acceptance use
Dynamic rejection rateFraction of dynamic/transient LiDAR or map points removed from the static map layerPrevent aircraft, GSE, and people from becoming localization map ghosts
Static preservation rateFraction of valid static infrastructure retainedProtect stand geometry, terminal edges, poles, markings, curbs, and docking features used by localization
FOD retention / alert rateFraction of FOD-like objects preserved as current hazards or surfaced as alerts rather than cleaned away as noiseDo not let map cleaning erase small safety-critical debris
Movable-static classification accuracyCorrectly classify parked aircraft/GSE, temporary barriers, cones, chocks, and staged carts as movable-static or review-requiredAvoid promoting temporary objects into the permanent map
False-free-space rateRate at which cleaned maps or occupancy outputs imply free space where a hazard existsCore safety metric for planner and safety-case review
Ghost rate per standRemaining transient dynamic points per stand, gate, route segment, or 100 mPractical map QA metric for apron operations
Localization deltaATE/RPE, scan-to-map residual, inlier ratio, degeneracy, and relocalization success before and after cleaningCleaning must not reduce localization integrity
Reviewer burdenAlerts or manual QA minutes per stand/km and percentage accepted by reviewersEnsures the fleet map workflow can scale
Operational latencyTime from capture to quarantine/update decision and time to publish a safe map packageSupports controlled AGVS test plans and map-change response

Airside/Indoor/Outdoor Transfer

Proxy dataWhat it can teachWhat must be collected airside
FOD-AFOD object categories, image detection under light and weather tags3D LiDAR/camera/radar FOD labels on actual apron concrete and taxiway surfaces
AIT ApronAirport-apron object taxonomy and environmental slicesFull sensor-suite logs, calibrated LiDAR/camera/radar, ego poses, and map-layer labels
KTH/SemanticKITTI-derived map-cleaning benchmarksPR/RR style map-cleaning metrics and reproducible cleaner comparisonsAircraft-present/absent route pairs, GSE staging changes, stand-closure changes, wet/night/de-icing captures
Indoor moved-object datasetsAdded/removed/moved object logic and object persistenceOutdoor geodetic alignment, GNSS/INS failure zones, and aircraft-scale occlusion
FAA AGVS guidanceControlled test envelope, human monitor, test-plan evidence, safety responsibilitiesTechnical pass/fail thresholds agreed with airport sponsor and safety authority

Validation Guidance

  1. Define map layers before testing: permanent static, movable-static, current dynamic, FOD/hazard, artifact, and unknown/review.
  2. Build paired captures for each stand: quiet survey, busy operation, aircraft present, aircraft absent, GSE staged, GSE removed, wet/dry, day/night, and representative weather.
  3. Run at least two independent cleaners and compare disagreement. Candidate baselines should include ERASOR, Removert, MapCleaner, and a MOS-pre-filtered variant.
  4. Preserve FOD and small hazards as current-world alerts even if they should not enter the permanent static map.
  5. Tie every benchmark run to a safety-case artifact: route, ODD, participants, human monitor assumption, takeover path, sensor configuration, RF status if relevant, and map version.
  6. Quarantine changed map tiles until cross-session evidence or human review confirms update, removal, or temporary override.
  7. Set stricter thresholds near aircraft movement, pedestrian zones, docking areas, and blind-corner service roads than in open apron transit zones.

Sources

Public research notes collected from public sources.