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."

1. What a LiDAR Point Measures ​

A scanning LiDAR point is usually 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

The spherical measurement model is:

z = [r, theta, phi, I]

x = r * cos(phi) * cos(theta)
y = r * cos(phi) * sin(theta)
z = r * sin(phi)

where r is range, theta is azimuth, phi is elevation, and I is an intensity-like return strength. Most AV LiDARs do not directly observe a perfect point on a surface; they observe a finite beam footprint and report a range extracted from the return waveform.

Sensor Model Impact ​

TaskWhy the model matters
PerceptionObject size, freespace, curb height, and small obstacle detections 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, low-return, or weather-corrupted points.
MappingStatic maps accumulate systematic beam and timing errors into blurred walls, doubled poles, and biased ground planes.
ValidationA detector should be tested against range, incidence angle, intensity, weather, and motion bins, not only aggregate precision/recall.

2. Time-of-Flight LiDAR ​

Pulsed time-of-flight (ToF) LiDAR estimates range from round-trip travel time:

r = c * delta_t / 2

where c is the speed of light and delta_t is the time between pulse emission and return detection.

Phase-shift ToF uses a modulated continuous signal and estimates phase delay:

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

Phase systems can be ambiguous beyond a modulation-dependent range and often use multiple modulation frequencies. Pulsed automotive spinning and solid-state LiDARs usually hide waveform details behind point-cloud outputs, but the underlying physics still appears as range quantization, multi-return behavior, minimum range, maximum range, and intensity dependence.

Range Noise Sources ​

Range noise combines random timing jitter, finite pulse width, detector noise, range quantization, surface effects, and atmospheric scatter:

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

Practical modeling:

  • Start with datasheet range accuracy as a lower bound, not a deployment covariance.
  • Increase sigma_r for low intensity, long range, high incidence angle, transparent/black/specular materials, rain, fog, and partial beam hits.
  • Model per-channel range bias separately from random noise for high-accuracy mapping.
  • Keep return type. First and last returns have different obstacle and surface semantics in vegetation, spray, snow, and fence-like objects.

3. FMCW and Coherent LiDAR ​

Frequency-modulated continuous-wave (FMCW) LiDAR transmits a chirped optical signal and mixes the received light with a local copy. The beat frequency encodes range. Because the receiver is coherent, Doppler can also be measured.

For an ideal linear chirp:

f_b = (2 * S * r) / c
r = c * f_b / (2 * S)

S = B / T_chirp
range_resolution = c / (2 * B)

For radial velocity:

f_d = 2 * v_r / lambda

FMCW LiDAR can reject incoherent sunlight and some interference better than direct-detection ToF, and it can provide line-of-sight velocity. Its practical noise model must include chirp nonlinearity, coherence length, phase noise, speckle, range-velocity coupling, and scan pattern timing.

AV Relevance ​

Velocity-bearing LiDAR points can help distinguish moving aircraft equipment, ground crew, rain/spray returns, and static map structure. The fusion stack should still treat FMCW velocity as line-of-sight, not full 3D velocity.

4. Angular Noise and Beam Geometry ​

Angular uncertainty often dominates lateral error at long range:

sigma_lateral ~= r * sigma_angle

For example:

r = 80 m
sigma_angle = 0.05 deg = 0.000873 rad
sigma_lateral ~= 0.07 m

A point covariance in the LiDAR frame can be approximated by propagating spherical noise through the Cartesian projection:

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

Practical notes:

  • Spinning multi-beam LiDARs have channel-specific vertical angles and azimuth/time offsets.
  • MEMS, flash, OPA, and oscillating mirror sensors have scan-pattern-dependent angular noise and rolling acquisition.
  • Beam divergence creates a footprint that grows with range. Thin poles, wires, cones, and FOD can be partial hits rather than clean surface samples.

5. Intensity and Reflectance ​

LiDAR intensity is not a universal reflectance value. It depends on emitted power, receiver gain, range, incidence angle, target material, wavelength, surface roughness, atmospheric transmission, and vendor signal processing.

A simplified radiometric relationship is:

I_meas proportional_to
  P_tx * rho_target * cos(alpha) * A_receiver * eta /
  (r^2 * beam_spreading * atmospheric_loss)

where alpha is incidence angle and rho_target is target reflectance at the LiDAR wavelength. Some workflows correct intensity using range and incidence angle before using it for lane markings, retroreflective signs, calibration targets, or map layers.

Noise-model guidance:

  • Do not compare raw intensity across sensors without calibration.
  • Low intensity is a warning for higher range variance and higher drop probability.
  • Retroreflectors can saturate receivers and bias range extraction.
  • Wet markings and metallic aircraft surfaces can produce specular returns, missing returns, or multipath-like artifacts.

6. Incidence Angle and Surface Effects ​

Incidence angle affects both return probability and range accuracy. At grazing angles, the beam footprint stretches across a surface, the return waveform widens, and the reported range may come from an edge, mixed pixel, or leading part of the footprint.

cos_incidence = abs(n_surface dot (-ray_direction))
grazing when cos_incidence -> 0

Practical weighting:

sigma_r_eff = sigma_r_base / max(cos_incidence, epsilon)
drop_prob increases as cos_incidence decreases

This is especially important for:

  • ground-plane extraction at long range
  • aircraft fuselage and service vehicles with curved or shiny panels
  • glass, wet concrete, paint, puddles, and black rubber
  • ICP scan matching against walls seen nearly parallel to the ray

7. Weather, Aerosols, and Ray Drop ​

Rain, fog, snow, dust, and de-icing spray affect LiDAR through attenuation, backscatter, and false near-field returns.

P_return = P_surface * exp(-2 * beta * r) + P_backscatter

where beta is an extinction coefficient. The two-way exponential term matters because light travels to the target and back.

Operational effects:

  • Rain and spray create sparse false points near the sensor and reduce useful range.
  • Fog reduces contrast and produces distributed backscatter that can erase weak far objects.
  • Snow can create high-reflectance particles and partial occlusion.
  • Dirty windows and water films add fixed-angle artifacts and global intensity loss.

For perception validation, report performance by weather intensity and range, not just by scenario. For mapping, do not accept weather-corrupted scans into the long-lived map without filtering and traceability.

8. Motion Distortion and Deskew ​

A spinning LiDAR frame is not captured at one instant. Points are acquired over the scan period:

p_map(t_i) = T_map_lidar(t_i) * p_lidar_i

If the vehicle turns or accelerates during a scan, treating the whole cloud as captured at t_frame bends walls, tilts poles, and biases scan matching.

Deskewing needs:

  • per-point or per-column time
  • sensor-to-IMU/body extrinsic calibration
  • high-rate ego-motion from IMU, wheel odometry, or fused odometry
  • correct clock synchronization

For low-speed airside vehicles, motion distortion can still matter near an aircraft during tight turns because angular motion produces centimeter-scale projection errors at docking distances.

9. Calibration Hooks for SLAM and Mapping ​

LiDAR calibration spans intrinsics, extrinsics, timing, and intensity.

CalibrationParametersFailure symptom
Intrinsic beam calibrationvertical angles, azimuth offsets, range offsets, channel timingwavy walls, ring seams, blurred poles
Extrinsic to base/IMU/camera/radarT_base_lidar, T_imu_lidar, T_camera_lidarcolored point clouds misalign, deskew residuals grow
Time offsetLiDAR clock to IMU/PTP/GNSS clockresiduals grow during turns but not static scenes
Intensity/radiometricrange and channel response correctionlane markings and reflectors inconsistent across passes

Mapping pipelines should preserve:

  • raw point time, ring, intensity, and return type
  • calibration artifact ID
  • weather and sensor health diagnostics
  • pose covariance and scan-matching residuals

Without these fields, a map cannot be audited when a wall is doubled or a clearance envelope near an aircraft stand is wrong.

10. Noise Models for Estimation ​

Point-Level Model ​

Use anisotropic covariance. A simple point residual covariance is:

Sigma_point =
  R_ray * diag(sigma_r^2, sigma_tangent1^2, sigma_tangent2^2) * R_ray^T

sigma_tangent ~= r * sigma_angle

Inflate for:

  • range > reliable_range
  • low intensity or saturation
  • high incidence angle
  • weather and window contamination
  • dynamic object probability
  • mixed pixels near depth discontinuities

Scan Matching Model ​

For ICP/GICP/NDT factors, the pose covariance should reflect geometry:

SceneConstraint quality
two perpendicular walls, poles, curbsstrong 6-DoF or strong planar constraint
long featureless wallweak along-wall translation
flat apron or open runwayweak yaw and horizontal translation without distant structure
mostly dynamic objectshigh outlier risk

Monitor the Hessian or information matrix from scan matching. Small eigenvalues indicate directions the scan does not constrain.

Practical Starting Point ​

clear weather point range sigma: 0.02 to 0.05 m
angular sigma: datasheet beam/encoder precision, often 0.02 to 0.1 deg
rain/fog/de-icing spray: inflate range and outlier probability by 2x to 10x
grazing incidence: inflate by 1 / max(cos_incidence, 0.2)

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

11. Failure Modes ​

Failure modeCauseMitigation
Ray drop interpreted as free spacedark material, rain, fog, glass, specular angletrack unknown space separately from observed free space
Doubled map structuretime offset, motion distortion, bad extrinsic, moving objectsdeskew, calibrate, dynamic filtering, map QA by pass
ICP converges to wrong poserepetitive geometry, open apron, poor initializationIMU/GNSS prior, robust kernels, degeneracy detection
Overconfident scan factorfixed covariance in weak geometryderive covariance from registration Hessian and scene quality
Ghost or mixed returnsglass, wet ground, metallic aircraft panelsmulti-return logic, intensity/range gating, temporal confirmation
Weather false obstaclesrain, snow, spray, dustparticle filters, near-field weather classifier, radar cross-check
Channel seam artifactsintrinsic beam error or thermal driftchannel calibration and residual monitoring

12. Sources ​

Public research notes collected from public sources.

Copyright © 2026 kvynlim