Industry Research

Appearance

LiDAR Working Principles and Noise Models ​

LiDAR Working Principles and Noise Models curated visual

Visual: LiDAR point formation diagram showing emitted pulse or chirp, time-of-flight/FMCW measurement, beam angle, reflectance, incidence angle, weather dropout, and range noise.

LiDAR turns emitted light into range, bearing, and sometimes reflectance or velocity measurements. For perception it is a geometric sensor. For SLAM and mapping it is a source of dense surface constraints. The useful model is not "a point cloud is truth"; it is "each returned point is a range-bearing measurement whose uncertainty depends on beam geometry, target material, incidence angle, atmosphere, timing, and calibration."

Why It Matters ​

Choice / EffectImpactRisk if ignored
ToF vs. FMCWFMCW adds per-point velocity; different noise floorTreating FMCW radial velocity as 3D velocity corrupts tracker
Wavelength (905 vs. 1550 nm)Eye-safety budget, fog/mist tolerance, solar noise905 nm degrades significantly in de-icing mist
Beam divergenceControls mixed-pixel rate at depth edgesWide-divergence sensors miss wires and apron markers at range
Incidence-angle dependenceReflectance and range noise scale with cos(alpha)Ground markings at grazing incidence return near-zero intensity
Raw vs. calibrated intensity~4% mIoU gain from calibrated reflectivityModels fail across passes, sensors, and temperatures
Weather attenuation2–10× outlier rate; reduced effective rangeClear-weather perception fails on foggy apron
Motion distortion50 cm smear at 5 m/s over 100 ms sweepPainted markings smear; scan matching diverges
Time synchronization1 ms LiDAR–IMU offset → cm-class deskew errorGhost duplicates in multi-LiDAR setups

1. What a LiDAR Point Measures ​

A scanning LiDAR point is stored as:

p_lidar = [x, y, z]
intensity = returned signal metric, vendor-specific
ring = laser channel or scan line
time = per-point or per-column acquisition time
return_type = strongest, first, last, dual, or vendor-specific
radial_velocity = FMCW sensors only; signed m/s

The spherical measurement model:

z = [r, theta, phi, I]
x = r * cos(phi) * cos(theta)
y = r * cos(phi) * sin(theta)
z = r * sin(phi)

The sensor measures natively in spherical coordinates; the Cartesian conversion introduces non-uniform spacing: tangential point spacing grows linearly with r, so ground-plane density (points/m²) falls as 1/r². Along the radial direction spacing is fixed by the pulse repetition interval.

Sensor Model Impact ​

TaskWhy the model matters
PerceptionObject size, freespace, curb height inherit range and angular noise. Ray drops can look like free space unless explicitly modeled.
SLAMICP/GICP residuals assume a point covariance. Wrong covariances over-weight grazing-angle and weather-corrupted points.
MappingStatic maps accumulate systematic beam and timing errors into blurred walls, doubled poles, and biased ground planes.
ValidationTest against range, incidence angle, intensity, weather, and motion bins — not only aggregate precision/recall.

2. Measurement Principles: ToF, AM-CW, and FMCW ​

2.1 Pulsed Direct Time-of-Flight (ToF) ​

Dominant architecture for automotive and mobile laser scanning sensors. A laser fires a short pulse (~few ns); the receiver measures the round-trip time:

r = c * delta_t / 2

where c = 3e8 m/s. Receivers use APD or SPAD detectors. Typical pulse widths 2–10 ns produce range ambiguity windows of 0.3–1.5 m, resolved by time-gating.

2.2 Amplitude-Modulated Continuous Wave (AM-CW) ​

The laser power is sinusoidally modulated at f_mod; range is from the phase shift of the returning envelope:

r = c * delta_phi / (4 * pi * f_mod)

Practical for short-range depth cameras; rarely used in long-range automotive class due to inherent range ambiguity beyond one modulation wavelength.

2.3 Frequency-Modulated Continuous Wave (FMCW) ​

FMCW adapts coherent radar to optical wavelengths. The transmitter emits a linearly chirped continuous beam; a local oscillator mixes with the return to produce a beat frequency proportional to range:

f_beat = kappa * tau_D     where tau_D = 2r/c
r = f_beat * c / (2 * kappa)
kappa = B / T_chirp                   (chirp rate, Hz/s)
range_resolution = c / (2 * B)
sigma_r ~= c / (2 * B * sqrt(SNR))   (Cramer-Rao precision)

Simultaneous velocity via Doppler (up-chirp / down-chirp pair separates range and velocity):

f_D = 2 * v_r / lambda

FMCW advantages: coherent detection rejects solar photons; lower peak power improves eye safety; per-point signed radial velocity enables dynamic/static separation without temporal differencing. Disadvantages: requires high-quality tunable lasers, high-speed ADC, FPGA-class FFT; currently higher cost than ToF. Treat FMCW velocity as line-of-sight (radial only), not full 3D velocity.

3. Sensor Architectures and Scan Patterns ​

3.1 Spinning Mechanical ​

A motor rotates N laser/detector pairs at 10–20 Hz. Each laser fires at a fixed elevation angle; the rotating platform sweeps 360° in azimuth, producing a cylindrical scan with uniform azimuth spacing (~0.1°–0.4°) and fixed, non-uniform elevation layers. One revolution = one scan (~50–100 ms at 10–20 Hz).

Point density is range-dependent in the horizontal plane (tangential spacing proportional to r) but the number of rings is fixed. 64-channel sensors (HDL-64E, OS1-64) provide 2–4× denser elevation coverage than 16-channel, substantially improving thin-class recall. Gaps between rings widen with range; thin horizontal features (wires, painted lines) can fall between rings at distance. Mature technology with high MLS deployment reliability; rotating seal limits IP-rating longevity.

3.2 MEMS ​

An oscillating mirror steers the beam with no full rotation. Scan patterns include raster, Lissajous, or pseudo-random over limited FOV (e.g., ±25° horizontal). Lissajous scan produces non-uniform density — denser at FOV centre/edges — requiring density-normalized processing. Currently the dominant commercial solid-state approach (Innoviz, Blickfeld, Hesai AT128).

3.3 OPA and Flash ​

OPA: phase-shift waveguide arrays steer the beam electronically — no moving parts. Currently limited FOV and power efficiency; promising for future sub-$100 costs.

Flash: a single large divergent pulse illuminates the entire scene; a 2D SPAD/APD array detects returns simultaneously. No motion distortion per flash; range limited by power/FOV trade-off. Suited for short-range proximity sensing (AGV docking, drone landing).

4. Wavelength Choice: 905 nm vs. 1550 nm ​

4.1 Eye Safety (IEC 60825-1 Class 1) ​

Below ~1400 nm the beam focuses on the retina; MPE is low. At 1550 nm water absorption in the cornea prevents retinal exposure; the Class 1 MPE is approximately 40× higher than at 905 nm, allowing far more emitted optical power for the same safety class.

4.2 Detectors and Solar Background ​

WavelengthDetectorCostSolar background
905 nmSilicon APD / SPADLowHigh (solar peak ~500–900 nm)
1550 nmInGaAs / Germanium APDHigherLow (solar falls steeply above 1400 nm)

1550 nm: better collimation (4× smaller spot at 100 m in direct comparison); lower solar background → improved daytime SNR; 20–40% practical range advantage in direct-sunlight conditions.

4.3 Weather Performance ​

  • Fog / mist: Mie scattering cross-section decreases with wavelength → 1550 nm has lower extinction. At 50 m visibility, attenuation ~9.2 dB; 1550 nm gives measurably longer effective range. Airport apron de-icing mist (propylene-glycol aerosol) acts as artificial fog — 1550 nm is preferable.
  • Rain: both wavelengths broadly comparable. Rain at 98 mm/h → ~22.7% FDR at 20 m; distance errors up to 4.9 cm at 20 m.
  • Snow: 905 nm may have slight edge in heavy snowfall; 1550 nm water- absorption resonance is a disadvantage in wet snow.

Practical airside recommendation: 1550 nm for apron deployments where de-icing mist and fog are primary risks. 905 nm adequate for clear-weather high-speed outdoor domains where cost matters.

5. Geometry of Measurement ​

5.1 Polar-Native Representation and 1/r² Density ​

The 1/r² density fall-off is fundamental: far more returns per unit area near the sensor than far away. Tangential spacing is fixed in angle but grows linearly with r; radial spacing is fixed by the pulse repetition interval. Aggregated multi-scan maps smooth the variation by accumulating returns from many poses, but residual density gradients affect segmentation model behavior.

5.2 Beam Divergence and Footprint ​

For beam divergence half-angle gamma, the footprint diameter at range r:

d_footprint ~= 2 * r * tan(gamma/2) ~= r * gamma    [small angle]

Examples: Teledyne Optech CL-360, gamma ~= 0.3 mrad → d = 3 cm at 100 m; typical automotive spinning sensor, gamma ~= 3 mrad → d = 30 cm at 100 m.

Larger footprint at range has two effects:

  1. Energy spreading: received power from a Lambertian target falls as 1/r² (footprint area ∝ r² compensates geometric spreading).
  2. Mixed pixels: footprint straddles depth discontinuities → flying points at incorrect intermediate ranges (§8.3).

5.3 Incidence Angle Effect ​

For a Lambertian surface, reflected intensity scales with cos(alpha) where alpha is the angle between beam and surface normal:

sigma_r_eff = sigma_r_base / max(cos(alpha), epsilon)
drop_prob increases as cos(alpha) -> 0

Grazing incidence causes low SNR, widened range noise, elongated footprint, and systematic range bias toward the leading edge of the footprint. Critical for: ground-plane extraction at range; aircraft fuselage; glass and wet concrete; apron markings at low sensor mounting height.

6. Intensity and Reflectance ​

6.1 Raw Intensity Model ​

The sensor reports a dimensionless integer proportional to returning pulse amplitude. The essential dependencies for calibration:

I(r, alpha, rho) ~= eta(r) * rho * cos(alpha) / r^2

Where rho = surface reflectivity (0–1 diffuse; > 1 retro-reflectors), alpha = incidence angle, r = range, eta(r) = near-range defocus correction. A white wall at 50 m returns ~4× less raw intensity than at 25 m; a surface tilted 60° returns half the power of a perpendicular one. Neither effect is a material property — both must be removed.

6.2 Calibrated Reflectivity ​

To obtain material-intrinsic reflectivity:

rho_cal = I(r, alpha) * r^2 / [cos(alpha) * eta(r)]

Field evidence (arXiv 2403.13188 "Reflectivity Is All You Need"): replacing raw intensity with rho_cal yields ~4% mIoU improvement on off-road segmentation (12 of 13 classes improved) and ~3.8% F1 improvement on road-marking detection across lighting conditions. Largest gains for material-based classes (ground cover, painted markings).

For multi-pass MLS maps: (1) apply r²/cos(alpha) correction per point before accumulation; (2) verify cross-pass consistency on stable targets; (3) average corrected reflectivity across voxel occupants — averaging suppresses shot noise and scan-to-scan variability; (4) report per-voxel standard deviation as a map quality metric.

6.3 Per-Laser Calibration ​

Multi-beam sensors use N distinct laser/detector pairs with different output power, responsivity, and optical-path efficiency. The Velodyne HDL-64E ships a 64-entry XML calibration file (rotational, vertical, distance, intensity corrections per channel). Without it, identical targets appear at different intensity values on different rings — a systematic artefact that corrupts intensity-based segmentation. Per-layer adaptive thresholding (Otsu per ring) is necessary precisely because of these inter-laser differences (arXiv 2211.01105). Calibration drift over temperature requires periodic recalibration.

6.4 Retro-Reflectors and Intensity Saturation ​

Retro-reflective materials (road-sign sheeting, high-vis vest strips, airport PAPI reflectors) have rho >> 1. Effects:

  • Saturation: ADC clips; intensity is maximum-pegged and meaningless.
  • Blooming: charge from saturated SPAD bleeds into adjacent channels → false nearby returns.
  • Ghost returns: multi-bounce off adjacent specular surfaces → incorrect ranges.

This is the high-vis vest problem for airside AV: ground crew vest saturates returns, obliterates body outline, generates ghosts. Ground crew detection must be cross-modal (LiDAR geometry + camera); models need saturation augmentation.

7. Multi-Echo / Multi-Return ​

A single pulse produces multiple returns when the footprint straddles objects at different ranges. The backend detects peaks in the returned waveform.

Return modeDescriptionPrimary use
First returnClosest detectable targetObstacle detection; foreground edge
Last returnFarthest detectable targetGround beneath vegetation; wire background
Strongest returnHighest-amplitude peakBest SNR; general detection
Dual / all returnsFirst and last recordedVegetation, wire, weather separation

Velodyne VLP-32C: first-or-strongest vs. dual-return firmware modes. Typical commercial spinning sensors support 2–5 returns per pulse.

Airside use cases: wire/cable detection (weak first return = cable, strong last = background); weather filtering (near-field weak first returns from rain/fog particles; strong last likely solid surface); vegetation penetration at perimeter for Digital Terrain Model.

8. Noise Sources ​

8.1 Range Noise (Gaussian Model) ​

Timing jitter (comparator noise, APD dark counts, photon shot noise) translates to range error. For a Lambertian target under good conditions:

r_meas = r_true + b_r(channel, temperature, range, intensity) + epsilon_r
epsilon_r ~ N(0, sigma_r^2)

Typical values: sigma_r ~= 1–3 cm at 30 m, rising to 5–10 cm at 100+ m. Model per-channel range bias separately from random noise for high-accuracy mapping. FMCW: sigma_r ~= c / (2*B*sqrt(SNR)) — mm-class at moderate range with wide-bandwidth chirps.

8.2 Angular Noise and Point Covariance ​

Angular uncertainty dominates lateral error at long range:

sigma_lateral ~= r * sigma_angle

Example: r = 80 m, sigma_angle = 0.05 deg = 0.000873 rad → sigma_lateral ~= 0.07 m.

Full anisotropic point covariance from spherical noise propagation:

Sigma_xyz = J_sph2cart * diag(sigma_r^2, sigma_theta^2, sigma_phi^2) * J_sph2cart^T

Inflate for: range > reliable_range; low intensity or saturation; high incidence angle; weather; mixed pixels near depth discontinuities; dynamic object probability.

8.3 Mixed Pixels (Depth Edge Artefact) ​

At depth discontinuities, a single footprint simultaneously illuminates foreground and background. Electronics may return an interpolated "flying point" at an incorrect range. Flying points cluster as halo artefacts at object edges. Severity scales with beam divergence. In aggregated MLS maps, consistent scan directions allow partial suppression. See also LiDAR Artifact Removal Techniques.

8.4 Outlier Rate and False Returns ​

Sources of spurious points: atmospheric particles (dust, pollen); solar cross-talk (especially 905 nm); multi-path from specular surfaces; electronic noise floor. Typical outlier rates: 0.1–1% in clean conditions; exceeds 20% at ranges below 20 m in fog or heavy dust.

8.5 Specular Surfaces ​

Wet asphalt, glass, and polished metal reflect the beam away from the sensor, producing zero or very weak return. Consequences: ground plane drops points in wet conditions; glass walls invisible; multi-bounce creates ghost points at incorrect ranges; wet markings fall below intensity-threshold detectors.

8.6 Ambient Light (Sunlight Interference) ​

Sunlight in the detector time gate raises the noise floor. 905 nm sensors are more susceptible (solar irradiance peaks near 900 nm). FMCW sensors are largely immune due to coherent detection filtering incoherent photons.

9. Weather Effects ​

9.1 Mie Scattering and Attenuation ​

Two-way propagation through an attenuating medium:

P_return = P_surface * exp(-2 * beta * r) + P_backscatter
sigma_ext,fog ~= 3 / V     [V = visibility in metres; simplified]

For rain, the Marshall-Palmer distribution governs extinction:

N(D) = N0 * exp(-Lambda * D)
Lambda = 4.1 * rainfall_rate^{-0.21}

9.2 Fog ​

Both attenuation (signal reduction) and backscatter (spurious near-range returns). At 50 m visibility: ~9.2 dB attenuation. FDR in dense fog exceeds 500% at short ranges. 1550 nm has lower extinction → longer effective range.

9.3 Rain ​

Rain at 98 mm/h → ~22.7% FDR at 20 m; distance errors up to 4.9 cm at 20 m. Both 905 nm and 1550 nm comparably affected.

9.4 Snow and De-Icing Mist ​

Falling snow: dense small-particle backscatter; weak, short-range, spatially clustered intensity signature. Wet snow worse for 1550 nm (water absorption). De-snowing algorithms (DSOR, LiSnowNet, SLiDE) exploit geometric and intensity signatures.

De-icing mist on airport aprons (propylene-glycol aerosol) acts as artificial fog. 905 nm especially affected. Intensity thresholds calibrated in clear conditions need adaptive recalibration under de-icing. Models trained on clear data degrade in mist; weather-augmented training data required.

10. Motion Distortion and Deskew ​

10.1 The Mechanism ​

A spinning LiDAR at 10 Hz sweeps 100 ms. At 5 m/s, the sensor translates 50 cm during the sweep — a substantial geometric error. Each azimuth sector is captured at a different sensor pose (directly analogous to rolling shutter). For perception: markings smear, walls bend. For scan matching: undeskewed clouds appear bent, biasing ICP. For mapping: the same structure is doubled across passes acquired at different speeds.

10.2 Deskewing Algorithm ​

p_map(t_i) = T_map_lidar(t_i) * p_lidar_i
  1. Record per-point timestamp t_i (Ouster/Velodyne/Livox all provide this).
  2. Integrate IMU angular velocity and acceleration to reconstruct pose T(t_i).
  3. Transform each point to the reference frame at sweep-start T(t_0).

Critical dependencies: LiDAR–IMU time sync (1 ms → cm-class error at 5 m/s); extrinsic calibration (lever arm < 1 cm, rotation < 0.1°); IMU bias in state. See also Rolling Shutter and LiDAR Deskew.

11. Timestamping and Synchronization ​

11.1 Per-Point Timestamps ​

Ouster OS-series and Velodyne sensors embed per-point nanosecond-precision timestamps in the UDP packet stream (Ouster: 10 ns resolution counter referenced to active clock source). This t channel enables deskewing (§10), sensor fusion, and multi-sensor phase locking.

11.2 Clock Sources ​

ModeMechanismAbsolute accuracyUse case
Internal oscillatorFree-running monotonic counterN/ARelative timing only
GPS / 1-PPSOpto-isolated pulse phase-locks sensor clock~1 µsOutdoor mapping
IEEE 1588 PTPSensor as PTP slave synced to network master~100 µs network; 10 ns counterVehicle LAN; multi-sensor sync

Ouster supports phase locking between sensors on the same PTP master (prevents inter-sensor beam interference). Key sync impacts: 1 ms deskew offset ~= 5 mm at 5 m/s; 10 ms camera–LiDAR offset → 5 cm projected misalignment; unsync'd multi-LiDAR creates ghost duplicates on moving objects; PTP timestamps enable GNSS trajectory stitching across passes. See also Sensor Calibration and Time Synchronization.

12. Calibration ​

12.1 Intrinsic (Per-Laser) ​

For spinning sensors, each laser/detector pair requires: rotational correction (azimuth offset), vertical correction (elevation offset), distance correction (constant range offset), and intensity correction (per-channel gain/offset). The Velodyne HDL-64E ships these as a 64-entry XML file. Without applying calibration, identical targets appear at different intensity values on different rings — corrupting intensity-based segmentation. Periodic recalibration against known planar targets is advised for intensity-critical applications.

Standard re-calibration: scan a planar wall at multiple ranges and orientations; least-squares fit per-laser parameters to minimise planar residuals.

12.2 Extrinsic (LiDAR–IMU, LiDAR–Camera) ​

LiDAR–IMU: 6-DOF rigid transform. Protocol: excite all rotational and translational DOF; joint optimisation of LiDAR odometry and IMU pre-integration (e.g., LI-Init, FAST-LIO). Required accuracy: lever arm < 1 cm, rotation < 0.1°. Requires rotational excitation (figure-8 or 3D spiral trajectory).

LiDAR–camera: overlap in FOV needed. Methods: checkerboard target (plane-fit in LiDAR + PnP in camera); trihedron (corner cube); targetless via natural edges (needs good initial estimate).

Calibration typeParametersFailure symptom
Intrinsic beamangles, azimuth offsets, range offsets, channel timingwavy walls, ring seams, blurred poles
Extrinsic to IMUT_imu_lidar, lever armdeskew residuals grow during turns only
Extrinsic to cameraT_camera_lidarcolored point clouds misalign; depth-edge projection artifacts
Time offsetLiDAR clock to IMU / PTP / GNSSresiduals grow in motion; vanish when static
Intensity / radiometricrange and channel response correctionmarkings inconsistent across passes and temperatures

Mapping pipelines should preserve: raw point time, ring, return type, intensity, calibration artifact ID, weather diagnostics, pose covariance and scan-matching residuals — needed to audit map errors at aircraft stand clearance zones.

13. Noise Models for Estimation ​

13.1 Point-Level Anisotropic Covariance ​

Sigma_point = R_ray * diag(sigma_r^2, sigma_t1^2, sigma_t2^2) * R_ray^T
sigma_tangent ~= r * sigma_angle

13.2 Scan Matching ​

SceneConstraint quality
Two perpendicular walls, poles, curbsStrong 6-DoF
Long featureless wallWeak along-wall translation
Flat apron or open runwayWeak yaw and horizontal translation
Mostly dynamic objectsHigh outlier risk

Monitor Hessian eigenvalues; small eigenvalues flag unconstrained directions — increase uncertainty in those directions rather than clamping to a minimum.

13.3 Practical Starting Values ​

clear weather point range sigma:    0.02 to 0.05 m
angular sigma:                      0.02 to 0.1 deg (datasheet)
rain / fog / de-icing spray:        inflate sigma and outlier prob 2x to 10x
grazing incidence (alpha > 60 deg): inflate by 1 / max(cos(alpha), 0.2)
retro-reflector zone:               discard intensity; flag range as possibly biased

Validate with residual histograms and NIS-like consistency checks against surveyed structures and repeated mapping passes.

14. Implications for Aggregated-Map Segmentation ​

14.1 Multi-Pass Intensity Calibration ​

In a multi-pass MLS map, the same ground patch is scanned from different ranges, incidence angles, and conditions (atmospheric, thermal). Production approach: (1) apply r²/cos(alpha) correction per point before accumulation; (2) verify cross-pass consistency on stable targets (concrete, asphalt); (3) average corrected reflectivity across voxel occupants; (4) report per-voxel standard deviation as quality flag. Result: voxel-averaged calibrated reflectivity approaches true material reflectance and enables stable class boundaries.

14.2 Density Signatures ​

  • Ring gaps widen with range; thin horizontal features fall between rings.
  • 64-channel sensors provide 2–4× denser elevation coverage than 16-channel, improving thin-class recall.
  • Overlap regions (multi-pass, reversed headings) create dense local distributions — subsample before training or they dominate gradients.
  • Evaluate mIoU stratified by local density (Low/Medium/High points/m²).

14.3 Sensor Choice → Segmentation Behaviour ​

FactorEffect on segmentation
High channel count (64+)Better thin-structure recall (wires, curbs, signs); finer vegetation/ground separation
1550 nmStable calibrated reflectivity in fog/mist → reliable class boundaries in adverse weather
905 nmSolar background spikes in low-angle sunlight → elevated false-return rate at horizon
FMCWPer-point velocity enables dynamic/static separation without temporal differencing
Motion-deskewed inputEliminates 50 cm marking smear at 5 m/s
Retro-reflector saturationClass-boundary failure near high-vis vests and sign retro-reflectors
Wide beam divergenceHigher flying-point rate at depth edges
Uncalibrated per-laser intensitySystematic inter-ring offset; requires per-ring normalization

14.4 Airside-Specific Considerations ​

  • Painted markings (taxiway centrelines, holding positions): rely entirely on calibrated reflectivity — geometry is near-flat. Otsu thresholding per ring is the production approach; requires per-laser calibration and r²/cos(alpha) correction beforehand (arXiv 2211.01105).
  • De-icing mist: propylene-glycol aerosol acts as artificial fog; intensity thresholds need adaptive recalibration. Consider 1550 nm; include weather-augmented training data.
  • High-vis vests: retro-reflective strips → saturation + ghost returns. Ground crew detection must be cross-modal; models need saturation augmentation.
  • Grazing incidence markings: sensor mounted low → apron markings nearly coplanar with beam at short range. Without incidence-angle correction, marking intensity varies with heading and vehicle speed.
  • The conditioning pipeline in Aggregated-Map Semantic Segmentation §9 addresses these sensor-physics artifacts at the data level.

15. Implementation Notes ​

  • Clip r²/cos(alpha) calibrated reflectivity at ~10× the 99th-percentile for Lambertian targets to prevent retro-reflector amplification to extreme values.
  • Preserve raw intensity, ring, time, and return type in the map store; calibrated reflectivity is derived — re-derivable with updated calibration.
  • Compute per-ring Otsu histograms across the full pass, not per scan; single-scan histograms may lack sufficient marking/non-marking balance.
  • Apply LiDAR–IMU extrinsic before integrating IMU increments; the lever arm amplifies angular velocity error if not applied at the IMU output stage.
  • Verify PTP phase locking is active in multi-LiDAR setups; unsync'd overlapping beams produce cross-talk returns in the merged cloud.
  • Below ~2 m range, the near-range defocus term eta(r) dominates; validate the near-range reflectivity model before using it for marking detection.
  • Apply voxel-grid subsampling before training on aggregated maps — dense near-field regions otherwise dominate gradients disproportionately.
  • Treat intensity-saturated points as a distinct bin; include saturated-return examples in every training batch for high-vis-vest and sign classes.

16. Failure Modes ​

SymptomCauseDiagnostic
Ray drop interpreted as free spaceDark material, rain, fog, glass, specular angleTrack hit/no-hit probability; separate unknown from observed free space
Doubled map structureTime offset, motion distortion, bad extrinsic, moving objectsDeskew; calibrate; dynamic filtering; per-pass map QA
ICP converges to wrong poseRepetitive geometry, open apron, poor initializationIMU/GNSS prior; robust kernels; Hessian eigenvalue degeneracy detection
Overconfident scan factorFixed covariance in weak geometryDerive covariance from registration Hessian and scene quality
Ghost or mixed returnsGlass, wet ground, aircraft panels, retro-reflectorsMulti-return logic; intensity/range gating; temporal confirmation
Weather false obstaclesRain, snow, de-icing sprayNear-field weather classifier; dual-return gating; radar cross-check
Channel seam artifactsIntrinsic beam error or thermal driftPer-channel calibration; residual monitoring over temperature range
Intensity inconsistent across passesMissing r²/cos(alpha) correction; thermal driftCalibrated reflectivity pipeline; cross-pass consistency check
Paint marking detector fails at distancePoints falling between rings; grazing low returnEvaluate per-ring, per-range-bin; check channel count vs. feature size
High-vis vest causes false geometryRetro-reflector saturation + bloomingFlag saturated points; cross-check with camera; add saturation augmentation
Deskew worsens accuracyLiDAR–IMU time offset or extrinsic error; IMU biasCheck residuals during rotation vs. translation; validate at known structure
Marking intensity varies with headingIncidence angle not correctedVerify r²/cos(alpha); validate same marking from multiple headings

17. Key Formulas ​

FormulaMeaning
r = c * delta_t / 2Pulsed ToF range
r = f_beat * c / (2 * kappa)FMCW range from beat frequency
delta_r = c / (2 * B)FMCW range resolution
sigma_r ~= c / (2 * B * sqrt(SNR))FMCW range precision
d_footprint ~= r * gammaBeam footprint diameter at range r
I(r, alpha, rho) ~= eta(r) * rho * cos(alpha) / r^2Intensity model
rho_cal = I * r^2 / [cos(alpha) * eta(r)]Calibrated reflectivity
sigma_ext,fog ~= 3 / VFog extinction from visibility V (m)
sigma_lateral ~= r * sigma_angleLateral position error from angular noise

18. Sources ​

Public research notes collected from public sources.

Copyright © 2026 kvynlim