Skip to content

Potentially Dynamic Object Removal by Ground Projection

Related docs: Potentially Dynamic Object Map Policy · Static-But-Transient Point Removal · LiDAR Map Cleaning and Dynamic Removal · ERASOR · Removert · Raymoval · FreeDOM · MapCleaner · BeautyMap · Aggregated-Map Semantic Segmentation · Map Construction Pipeline

Last updated: 2026-05-24


What It Is

"No More Potentially Dynamic Objects" (Woo, Jung, Kim, arXiv 2407.01073) is a static point-cloud map generation pipeline that removes not only currently moving objects but also objects that are stationary during mapping yet belong to a movable class. The method uses LiDAR 3D object detection to find potentially dynamic objects, ground segmentation to recover traversable surface context, and ground projection to replace object points with ground points before stacking the map.

The paper targets the common failure of geometry-only map cleaners: a parked vehicle, staged cart, or stationary person can remain still for the entire survey and therefore never creates the free-space contradiction that ERASOR, Removert, MapCleaner, FreeDOM, DUFOMap, or STATIC-LIO need. A detector can still identify the object class and route the points away from the permanent map.

Core Pipeline

text
LiDAR frames
  -> 3D object detector
  -> movable-class boxes
  -> ground segmentation
  -> object-point projection to ground
  -> SLAM / mapping on projected frames
  -> static point-cloud map + rejected-object evidence

The public implementation describes a concrete road-AV stack:

  • KITTI-format LiDAR input.
  • VoxelNeXt as the 3D object detector.
  • Patchwork++ for ground segmentation.
  • Ground-projection scripts for detected object regions.
  • SC-A-LOAM for mapping projected point clouds.

Those implementation choices are not the only possible stack. In an AV or non-road urban deployment, the detector could be CenterPoint, TransFusion, PV-RCNN, OpenPCDet, a camera-LiDAR detector, or a domain-specific apron/yard detector. The architectural contract is the important part: movable-class object geometry is withheld from the permanent map before accumulation, and the ground continuity beneath it is reconstructed conservatively.

Why It Is Different from Classical Dynamic Removal

Method familyEvidence typeHandles moving residuals?Handles parked-but-movable objects?
ERASOR / ERASOR++Pseudo-occupancy drop and ground restorationYesNo, unless the object later leaves during the observation window.
Removert / RaymovalRange-image ray contradictionYesNo, unless free-space contradiction appears.
FreeDOM / DUFOMapConservative free-space or void evidenceYesNo, if the object is never observed absent.
MapCleaner / BeautyMapTerrain, observation voting, or binary matrix comparisonYesLimited; stationary movable objects can vote as static.
Detector + ground projectionSemantic object class plus local ground modelYes, if detector sees the classYes, for classes covered by detector/taxonomy.
Lifelong map version controlMulti-session positive/negative changesYes after multiple passesYes after multi-session evidence or policy quarantine.

Detector-ground-projection methods are therefore best seen as the single-survey semantic branch of static-but-transient removal. They are complementary to multi-session methods: the detector quarantines movable-class objects before the second survey exists, while lifelong version control confirms or rejects changes across time.

Inputs and Outputs

InterfaceRequired fieldsNotes
LiDAR framesRaw point clouds, timestamps, calibration, ego poses or SLAM front-end inputWorks before or during accumulation.
Object detections3D boxes, class, confidence, detector version, class taxonomyClass policy determines which boxes are removable.
Ground segmentationGround/non-ground mask, plane or patch model, quality flagsNeeded to avoid holes after object removal.
Projection outputModified frame with object points projected to estimated groundShould be tagged as synthetic/conditioned geometry.
Rejected evidenceOriginal object points, boxes, classes, confidence, frame IDsRequired for audit, FOD review, and false-deletion debugging.
Static mapAccumulated map from projected framesInput to localization, semantic segmentation, or annotation.

Ground-Projection Contract

Removing a parked object from a mapping scan creates two risks:

  1. The object footprint can become a hole or false obstacle in the ground surface.
  2. Projecting the object points to ground can create false evidence that the space is currently clear.

The safe contract is:

  • Projection fills map geometry only for the static-map product.
  • The projected ground is not treated as live free-space evidence.
  • Original object points remain available in a rejected-object layer.
  • The map manifest records the detector class, confidence, box geometry, ground model, and projection method.
  • FOD-like small objects and unknown objects are routed to review instead of being projected away.

This is especially important for airside and industrial yards. A chock, tool, cone, cable, or dropped debris item might be small and removable, but it is also safety-relevant. It should not be erased by a generic movable-object cleaner.

Rejected-Object Evidence Sidecar

A production implementation should emit a rejected-object sidecar for every detection-driven exclusion or projection. That sidecar is the single-survey counterpart to the static-transient release-state matrix.

FieldReason
object_id, frame_ids, tile_ids, source_point_digestReconstructs the exact points that were excluded or projected
detector_model_id, weights_digest, class_id, class_confidence, box_digestMakes detector evidence reproducible and debuggable
ground_model_id, ground_quality_flag, projection_method, projected_point_digestSeparates synthetic ground continuity from original observed geometry
policy_actionOne of project_to_ground, exclude, movable_static, static_transient, fod_candidate, artifact, or unknown_review
reason_codeExplains the decision, e.g. movable_asset_quarantine, human_exclusion, small_ground_unknown, or zone_policy_override
review_state, waiver_id, source_map_acceptance_package_idLinks the object to publication decisions and source-map acceptance

The semantic-map manifest should reference aggregate digests derived from these sidecars through outputs.map_hygiene_layer_digests. Without the sidecar, the system can show that points disappeared, but not whether they disappeared because they were a person, a belt loader, FOD, a detector false positive, or weak ground projection.

Class Policy

Class familyDefault action in static-map buildReason
Cars, vans, trucks, busesExclude or project ground under footprintMovable road objects are poor static anchors.
Pedestrians, cyclists, workersExclude; do not create permanent geometrySafety actor and not map structure.
Aircraft and large GSEExclude from permanent layer; keep current-occupancy layerCan remain stationary for hours but should not anchor long-term maps.
Cones, barriers, temporary signsQuarantine unless work-order policy says otherwiseMay be temporary restriction, not permanent infrastructure.
Chocks, tools, debris, FODReview/hazard layer; do not silently project awaySmall hazards must remain actionable.
Fixed poles, walls, curbs, buildingsPreservePermanent localization structure.
VegetationPolicy dependentTrees can be long-lived; branches/leaves can change seasonally.
Unknown objectQuarantine or reviewUnknown does not mean safe to erase.

Failure Modes

Failure modeEffectMitigation
Detector false negativeParked/movable object remains in the permanent mapUse multi-session differencing, semantic segmentation, and targeted domain classes.
Detector false positiveFixed infrastructure is removed or projected to groundRequire class-specific geometry checks and static-feature preservation tests.
Box leakageParts of the object remain outside the detection boxDilate boxes cautiously, cluster residuals, and inspect rejected evidence.
Ground segmentation errorProjection creates ramps, holes, or false surfacesUse ground-quality flags, multi-frame smoothing, and no-publish thresholds for weak ground models.
Taxonomy mismatchRoad classes miss GSE, aircraft, cones, chocks, pallets, cartsTrain domain classes and version the class policy with the map manifest.
False-free-space interpretationPlanner assumes a live obstruction is absentKeep projected geometry isolated from runtime occupancy and FOD clearance.
Domain shiftKITTI-trained detector underperforms in depots, campuses, aprons, or night/weatherUse domain data, active learning, confidence calibration, and review queues.
Localization regressionRemoving large parked objects reduces scan-to-map constraintsRun localization A/B tests and preserve enough fixed infrastructure around object footprints.

Domain Fit

DomainFitRequired adaptation
Urban road AVStrongMature vehicle/person detector classes and road-ground priors.
Campus or plaza mappingConditionalAdd classes for benches, kiosks, crowds, bicycles, vendor stalls, and temporary fences.
Airport airsideStrong architecture, high taxonomy burdenAdd aircraft, belt loaders, tugs, carts, dollies, stairs, cones, chocks, and FOD policy.
Logistics yards and portsStrongAdd trailers, containers, forklifts, pallets, and blocked-lane policy.
Warehouse indoorConditionalGround projection may fail under racks, mezzanines, and stacked pallets.
Construction/miningConditionalMany "objects" are actually changing terrain or valid temporary works; policy review is mandatory.

Where It Sits in an End-to-End Semantic Map Pipeline

Detector-ground-projection should run before aggregated-map semantic segmentation:

text
raw LiDAR/RGB survey
  -> synchronization, calibration, SLAM
  -> dynamic residual cleaning
  -> detector-based movable-class quarantine/projection
  -> static candidate map
  -> semantic segmentation and label refinement
  -> publication gate with rejected-object evidence

The segmentation model then sees a cleaner static candidate map, while the publication gate still has access to the original rejected objects. This avoids a common failure where a segmentation model learns parked cars, staged GSE, or stationary people as permanent map structure.

Implementation Notes

  1. Treat the detector as a policy input, not an oracle. A class and confidence should route points to permanent, transient-candidate, FOD/hazard, or reviewer states.
  2. Keep original object points in object-space and map-space coordinates. Reviewers need to see what was removed and where it came from.
  3. Log detector model version, training domain, class set, and calibration state. A taxonomy change can invalidate old map-cleaning behavior.
  4. Separate static-map projection from runtime occupancy. A projected ground patch is useful for map continuity, but it is not proof that the area is clear now.
  5. Validate on object footprints and behind-object infrastructure. The cleaner must preserve fixed poles, walls, signs, curb edges, jet-bridge geometry, terminal edges, and other static anchors near parked movable objects.
  6. Use multi-session version control as a backstop. Detector-based removal is strongest before the second pass exists; PD/ND lifecycle evidence is stronger after repeated surveys.
  7. Export rejected-object sidecars into the source-map acceptance package and semantic-map hygiene layers. If a detection-based projection is not traceable, publication should treat it as a waiver, not as clean evidence.

Validation Gates

GatePass condition
Movable-object removalMovable-class objects are absent from the permanent map or routed to quarantine.
Static preservationFixed infrastructure near detections is preserved within tolerance.
Ground continuityObject footprints do not create holes, bumps, or false walls in the static candidate map.
Hazard retentionFOD-like or unknown objects are retained in review/hazard outputs.
Localization regressionScan-to-map localization error does not worsen after removal.
Evidence auditEvery removed/projected cluster can be traced to detector, ground model, and frame IDs.
Hygiene-layer digestRejected/projected objects appear in the semantic-map movable-static, static-transient, FOD-candidate, artifact, or unknown-review layer digests.
Domain-slice testNight, rain, glare, sparse returns, dense clutter, and non-road classes have separate metrics.

Sources

Public research notes collected from public sources.