Potentially Dynamic Object Map Policy

Last updated: 2026-05-24

Why It Matters

Potentially dynamic objects are objects that may be stationary now but can move later. In road mapping this includes parked vehicles. In airside mapping it includes aircraft, tugs, carts, stairs, loaders, buses, cones, chocks, dollies, FOD equipment, and temporary barriers.

The policy goal is not simply "remove dynamic objects." It is to prevent movable objects from contaminating the permanent static map while still preserving current hazards and operational blockages for the planner and safety system.

Source Lessons

Source	Technical lesson	Policy lesson
No More Potentially Dynamic Objects	Use 3D object detection, ground segmentation, and projection to build static point-cloud maps without parked vehicles	Object classes that can move should not become static anchors
YOLO-based potential dynamic removal	Camera detection plus LiDAR/NDT mapping can improve localization in traffic-heavy scenes	Multi-modal semantic filtering can outperform raw LiDAR segmentation alone
Dynamic map cleaning benchmark	Cleaning must preserve static infrastructure and FOD/hazard awareness	Map cleaning is a safety-validation problem
FAA FOD guidance	FOD is safety-critical when in an inappropriate location	Do not erase small hazards as "noise"
AIT Apron Dataset	Apron object detection needs domain-specific classes and environmental slices	Road vehicle classes are insufficient for airport policy

Policy Matrix

Class	Potentially dynamic?	Permanent static map	Live map / planner	Notes
Building, terminal edge, fixed pole	No	Keep/promote after QA	Hard obstacle	Primary localization candidates
Painted stand/service-road marking	No, but can be repainted	Keep/update after review	Route and alignment cue	Needs visibility/weather checks
Parked aircraft	Yes	Exclude	Occupancy obstacle and stand state	Never use as permanent localization anchor
GSE vehicle or cart	Yes	Exclude	Obstacle or route blockage	May be movable-static if parked
Cone/barrier line	Yes, unless work-zone approved	Exclude or temporary restriction	Block/slow/re-route	Promote only with ops confirmation
FOD/debris/tool	Yes/hazard	Exclude	Alert, stop, or avoid	Must not be cleaned away silently
Pedestrian/worker	Yes	Exclude	Dynamic actor	Person safety dominates map update
Snow/rain spray/reflection	Artifact/condition	Exclude	Condition-dependent caution	Keep evidence for QA if persistent

Processing Pipeline

Step	Action	Failure to avoid
Semantic detection	Detect known movable classes in camera/LiDAR	Treating parked objects as walls
Ground reasoning	Segment ground and project/remove object points where appropriate	Creating holes in navigable ground
Static preservation	Preserve fixed infrastructure behind/around removed objects	Removing terminal edges or poles with object masks
Hazard branch	Route FOD-like small objects to live alert workflow	Erasing safety hazards as map noise
Persistence tracking	Count sightings by tile, object class, and pose quality	Promoting one-pass observations
QA gate	Review candidate static changes before publication	Automatic deletion under occlusion
Regression	Run localization and route tests on updated tile	Cleaner improves appearance but harms localization

Permanence Evidence Ladder

Potentially dynamic classes need an evidence ladder because "stationary in the survey" is weaker than "belongs in the permanent map." The same object cluster can move through different release states as evidence accumulates:

Evidence stage	Typical evidence	Allowed layer	Promotion rule
Observed once	One pass, one source session, no operational record	Candidate or unknown-review	Never promote to permanent static
Stationary in capture window	Low scene flow/MOS, no free-space contradiction	Movable-static or static-transient	Use for current occupancy only; block training export as permanent truth
Recurrent but movable class	K-of-N sightings of aircraft, GSE, pallet, barrier, cone line, or parked vehicle	Movable-static, temporary overlay, or ops-review	Requires operations approval and expiry if it affects routing; default remains non-permanent
Recurrent fixed class	Multi-session fixed class with pose stability and localization utility	Permanent-static candidate	Promote only after class-preservation and localization replay pass
Operations-confirmed temporary	Work order, closure, construction plan, stand state, or maintenance schedule	Temporary overlay	Must include owner, start/end time, affected routes, and rollback rule
Safety/hazard candidate	Small object or debris in FOD-sensitive, stand, route, or pedestrian area	FOD/hazard or unknown-review	Never delete silently; reviewer or inspection workflow resolves
Artifact confirmed	Reflection, snow/rain spray, sensor bloom, stale duplicate, or map-registration error	Artifact/rejected evidence	Delete from base map but retain reason-coded evidence for audit

This ladder is also a training-data rule. Only permanent-static and reviewer-approved fixed-class points should become positive labels for base-map segmentation. Movable-static, temporary, FOD, artifact, and unknown-review points are valid supervision for quarantine and rejection heads, but they should not teach the model that a transient object is infrastructure.

Airside Class Policy

Object family	Default layer	Static promotion condition
Aircraft	Movable-static/current occupancy	Never, unless modeling a fixed static mockup or permanently installed exhibit outside ODD
Tugs, belt loaders, buses, carts	Movable-static/current occupancy	Never by default
Passenger stairs, dollies, ULD racks	Movable-static	Only if airport defines fixed storage infrastructure, not individual movable assets
Cones, temporary signs, portable barriers	Temporary restriction or movable-static	Work order/ops approval plus schedule
Chocks, tools, debris	FOD/hazard	Never; create inspection/removal event
Fixed signs, lights, poles	Permanent static	Multi-session observation and map QA
Pavement markings	Permanent static/regulatory	Multi-view evidence and approval when safety-critical

Relationship to Aggregated-Map Semantic Segmentation

This page defines the policy — which object classes may enter the permanent map and under what evidence. The end-to-end aggregated-map semantic segmentation pipeline (../../perception/overview/aggregated-map-semantic-segmentation.md) is one mechanism that operationalizes it.

• That pipeline's airside class taxonomy includes an explicit "staged GSE / permitted-static" class (its §6.3, class ID 9) — a deliberate quarantine class for equipment that is static now but is not permanent map structure. It is the learned-model realization of the movable-static layer in the Policy Matrix above.
• Division of labour: the segmentation model decides the class of every map point; this policy decides what happens to each class — exclude from the permanent map, keep as a live obstacle, route to a FOD alert. A learned per-point class (with confidence) plus this policy matrix yields an auditable, per-point layer assignment.
• Keeping the segmentation taxonomy aligned one-to-one with this policy matrix is what makes the learned output directly consumable: every class the model can predict must have a row here, and every movable class here must be a class the model is trained to recognize.
• The pipeline also enforces the FOD safety rule of the Source Lessons — it carries an explicit "unknown" catch-all class so small hazards are flagged for review, never silently cleaned away.

Cross-reference: ../../perception/overview/aggregated-map-semantic-segmentation.md §6.3 (taxonomy), §2.4 and §9.1 (dynamic removal vs. quarantine).

Acceptance Metrics

Metric	Target behavior
Movable-object rejection	High removal from permanent map
Static-feature preservation	No loss of fixed features behind movable objects
Ground continuity	Removed object footprints do not create false obstacles or holes
Hazard retention	FOD-like items appear as alerts, not map features and not discarded noise
Localization delta	Map cleaning improves or preserves scan-to-map localization
False-free-space rate	No claimed free space where a hazard or blocked route exists
Reviewer burden	Candidate static updates remain actionable per shift/stand

Implementation Guidance

1. Maintain two outputs from dynamic removal: a cleaned static candidate and a rejected-object evidence layer.
2. Use semantic class policy before geometry policy. A parked aircraft should be excluded even if it is observed consistently for many hours.
3. For 3D object detection methods, retain ground projection carefully so the map remains traversable but not falsely clear of live obstacles.
4. For camera-guided YOLO removal, log lighting/weather confidence because apron perception degrades at night, in glare, rain, fog, and de-icing conditions.
5. Never use a potentially dynamic removal model as the sole source of FOD clearance.
6. Tie all permanent-map promotions to source evidence, persistence, reviewer approval, and localization regression.
7. Before changing removal or promotion rules for potentially dynamic classes, rerun false-deletion validation on the local do-not-delete/adverse-airside holdout; unknown sparse clusters in route, stand, pedestrian, or FOD-sensitive zones route to review, not deletion.

Sources

• No More Potentially Dynamic Objects arXiv: https://arxiv.org/abs/2407.01073
• Local method page: Potentially Dynamic Object Removal by Ground Projection
• Enhancing LiDAR Mapping with YOLO-Based Potential Dynamic Object Removal: https://www.mdpi.com/1424-8220/24/23/7578
• FAA Foreign Object Debris Program: https://www.faa.gov/airports/airport_safety/fod
• FAA AC 150/5210-24A document page: https://www.faa.gov/airports/resources/advisory_circulars/index.cfm/go/document.current/documentNumber/150_5210-24
• AIT Apron Dataset: https://publications.ait.ac.at/de/datasets/apron-dataset/
• Local context: Movable-Static Layering for Airside Maps
• Local context: Airside Dynamic Map Cleaning Benchmark