LiDAR Ghost and Multipath Artifacts

Executive Summary

LiDAR ghost and multipath artifacts are false or distorted points caused by reflective surfaces, transparent materials, wet ground, high-intensity targets, receiver saturation, edge discontinuities, or external optical interference. They matter because they can create false objects, erase real surfaces, pollute static maps, and corrupt scan matching. On airside vehicles, these artifacts are common around wet aprons, terminal glass, polished aircraft skin, retroreflective markings, signs, cones, and high-visibility clothing.

Ghost handling belongs in the sensor layer and must feed perception, localization, mapping, and safety monitoring. It is not just a point-cloud cleanup step.

Artifact Taxonomy

Artifact	Mechanism	Point-cloud symptom	Airside examples
Multipath ghost	Laser reflects through one or more specular paths before returning	Object appears behind glass, below ground, or mirrored from a reflective plane	Terminal glass, wet apron, polished fuselage, service building windows.
Reflective dropout	Beam reflects away from receiver	Holes in expected surfaces	Standing water, glass, glossy painted markings.
Retroreflector blooming	Very high return overloads receiver or adjacent processing	Inflated target extent or nearby false points	Retroreflective signs, cones, vests, stand markings, vehicle plates.
Receiver saturation	Strong light or return exceeds useful measurement range	Angular dropout, clipped intensity, false range	Low sun, reflective aircraft, close reflectors.
Edge shadow points	Mixed returns near discontinuities	Ghost points near object edges	Aircraft gear, tow bars, poles, carts, fences.
Wet-surface mirror	Ground reflection creates indirect path	Below-ground points, mirrored obstacle fragments, missing ground	Rain puddles, wet concrete, de-icing fluid on apron.
Inter-sensor or external optical interference	Another emitter contaminates receiver timing or intensity	Sparse unexplained points or sector noise	Dense autonomous fleet, survey LiDARs, active optical beacons.

Detection Signals

Signal	What to check	Strength
Ray consistency	Does a point sit behind a known reflective surface or outside a feasible ray path?	Strong for maps with glass/wet-surface priors.
Ground consistency	Is a point below a surveyed or estimated ground plane?	Strong for wet-surface multipath.
Intensity saturation	Are intensity or reflectivity values clipped or extreme?	Strong for retroreflector bloom but hardware-specific.
Sector coverage	Are returns missing or corrupted in an angular sector?	Strong for blockage, sun, and receiver issues.
Temporal stability	Does the point persist as the ego pose changes?	Strong for moving ghosts; weak for persistent glass ghosts.
Multi-return/waveform	Are there multiple peaks or abnormal pulse shapes?	Strong where hardware exposes the data.
Cross-sensor agreement	Do radar, camera, thermal, or map evidence support the point?	Critical for safety decisions.

Removal and Mitigation Techniques

Technique	Best fit	Limitation
PCL ShadowPoints or edge-aware filtering	Edge ghost points around discontinuities	Does not solve glass or wet-ground multipath.
Reflective-plane modeling	Glass, mirrors, wet surfaces with known planes	Needs geometry and may miss curved aircraft surfaces.
Full-waveform ghost classification	Multi-path ghosts where waveform LiDAR is available	Not available on many automotive LiDARs.
Ground-model rejection	Below-ground wet-surface returns	Can delete valid curb/ramp/grade transitions if map is wrong.
Temporal occupancy and ray clearing	Ghosts that do not persist across viewpoints	Can remove thin valid structure.
Retroreflector extent gating	Bloom around known reflective signs or cones	Needs known size/shape priors and intensity calibration.
Sensor health diagnostics	Blockage, dust, sun sectors, saturation	Detects degradation but does not classify every ghost point.
Radar/camera/thermal confirmation	Safety-critical obstacle decisions	Each modality has its own failure modes.

Deployment Decision Rules

Condition	Action
Ghost point is outside physically plausible space and unsupported by other sensors	Remove or downweight; log as ghost candidate.
Ghost point could be a small obstacle in the near field	Keep as low-confidence obstacle unless radar/camera/thermal or temporal evidence rejects it.
Reflective artifact degrades localization map matching	Exclude from scan-to-map residuals and add a map-quality annotation.
Retroreflective bloom expands a known sign/cone/vest	Clip object extent using geometry, not intensity footprint alone.
Sector saturation or blockage is detected	Trigger sensor degradation state; do not treat denoising as recovery.
Wet apron creates below-ground returns	Use ground-model and cross-sensor checks; avoid updating persistent maps with affected points.

Failure Modes

• Ghost is accepted as a real obstacle, causing unnecessary braking or route blockage.
• Real person or chock is rejected as a ghost because it is sparse, low, or near a reflective surface.
• Static map accumulates mirrored structures, creating persistent localization bias.
• Scan matching aligns to a reflective artifact or ghost vehicle.
• Bloom around high-visibility clothing changes perceived pedestrian size.
• Filtered point cloud hides receiver saturation that should have triggered an ODD restriction.

Airside-Specific Validation Guidance

Build a test matrix around:

• Wet concrete and standing water at shallow scan angles.
• Retroreflective cones, signs, vests, stand markings, and vehicle plates.
• Aircraft fuselage and engine nacelle reflections.
• Terminal glass and jet bridge glass.
• Low sun with wet apron reflections.
• Night floodlights and reflective markings.
• De-icing fluid film and steam near aircraft.

Metrics:

• Ghost precision/recall by artifact type.
• False deletion rate for people, cones, chocks, tow bars, and aircraft gear.
• Detector false positives behind glass or below ground.
• Localization residual changes with reflective regions masked/unmasked.
• Static map ghost rate after repeated sessions.
• Sensor health response latency for saturation and blockage.

Implementation Notes

• Treat ghost removal as a candidate mask, not an irreversible delete, until safety validation is complete.
• Preserve raw, filtered, and ghost-candidate clouds in logs.
• Maintain per-sensor intensity calibration and saturation thresholds.
• Add map annotations for known glass, reflective signs, and wet-prone apron zones.
• Use radar as an independent confirmation source in fog, mist, spray, and wet-surface cases, while remembering radar also has multipath.
• Do not use a ghost filter to compensate for poor sensor placement; add near-field coverage or redundant sensors where required.

Sources

• Ghost-FWL project page: https://keio-csg.github.io/Ghost-FWL/
• PCL filters module, including ShadowPoints: https://pointclouds.org/documentation/group__filters.html
• PCL ShadowPoints reference: https://docs.ros.org/hydro/api/pcl/html/classpcl_1_1ShadowPoints.html
• Ill-reflecting surface robustness study: https://arxiv.org/abs/2309.10504
• Reflective noise filtering: https://www.mdpi.com/2072-4292/13/16/3058
• Autoware blockage diagnostics: https://autowarefoundation.github.io/autoware_universe/pr-10077/sensing/autoware_pointcloud_preprocessor/docs/blockage-diag/