Industry Research

Appearance

Thermal IR Radiometry First Principles ​

Thermal IR Radiometry First Principles curated visual

Visual: thermal IR radiometry measurement chain showing scene emission, reflected background, atmospheric path, optics, microbolometer response, NUC tables, radiometric parameters, and fusion-ready confidence.

Thermal infrared cameras do not measure object class or true temperature directly. They measure infrared radiance at the detector, then convert that signal into image intensity or apparent temperature using calibration tables, scene parameters, and assumptions about emissivity, reflection, atmosphere, optics, and camera drift.

The useful autonomy contract is therefore: "this pixel or feature is a radiance-derived measurement with known parameters, uncertainty, health state, and validity limits," not simply "thermal sees heat."

Why it matters for AV, perception, SLAM, and mapping ​

Thermal IR is valuable because it observes emitted long-wave or mid-wave radiation rather than reflected visible light. It can improve night, smoke, fog, dust, glare, and low-illumination operation. It can also fail quietly when the stack treats apparent temperature or thermal contrast as a stable semantic fact.

Radiometry matters when a system uses thermal data for:

  • night pedestrian, vehicle, animal, and equipment detection
  • hot engine, brake, tire, battery, motor, or fire monitoring
  • jet-blast, exhaust, and industrial heat-source awareness
  • thermal-inertial odometry, RGB-thermal matching, or thermal map layers
  • degraded-mode confidence, sensor health monitoring, and validation evidence

For autonomy review, the core question is whether thermal outputs are carried with enough context for downstream modules to know when they are trustworthy.

Measurement contract ​

Typical inputs:

  • raw or processed thermal frame, bit depth, frame time, and exposure/integration state
  • spectral band, lens/window material, optics transmission, and detector type
  • camera internal temperature, NUC/FFC state, bad-pixel map, and calibration table ID
  • object distance, emissivity assumption, reflected apparent temperature, atmospheric temperature, humidity, and path length where radiometry is used
  • camera intrinsics, distortion, extrinsics, timestamp offset, and synchronization source
  • optional reference target, blackbody check, or fleet health baseline

Typical outputs:

  • thermal intensity, radiance estimate, or apparent temperature per pixel
  • feature, detection, track, free-space cue, or thermal residual
  • validity and health flags: NUC event, saturation, low contrast, dead pixels, lens obstruction, stale frame, bad radiometric parameters
  • confidence, covariance, or quality score for fusion
  • provenance: radiometric versus contrast-only mode, parameter set, calibration version, and preprocessing path

The downstream consumer should know whether it is receiving true radiometric temperature estimates, relative thermal contrast, 8-bit display imagery, or a learned feature derived from a normalized frame.

Radiance model ​

For an opaque surface observed by a thermal camera, a compact radiometry model is:

text
L_sensor =
  tau_atm * [epsilon * L_bb(T_object)
             + (1 - epsilon) * L_reflected]
  + (1 - tau_atm) * L_atmosphere
  + L_optics
  + sensor_offset
  + noise

where:

  • epsilon is target emissivity
  • L_bb(T_object) is blackbody radiance at object temperature
  • L_reflected is reflected apparent radiance from surroundings
  • tau_atm is path transmission through air, humidity, smoke, fog, or dust
  • L_atmosphere is radiance emitted by the air path
  • L_optics includes lens, window, housing, and internal camera contributions

FLIR thermography documentation frames accurate temperature measurement around object parameters such as emissivity, reflected apparent temperature, distance, relative humidity, and atmospheric temperature. ASTM E1933 similarly treats emissivity as a field or laboratory compensation problem for infrared imaging radiometers.

For AV perception, many tasks need only stable contrast or feature evidence. For thermal safety claims, asset-temperature checks, fire detection, battery monitoring, and thermal maps, radiometric parameters and their uncertainty must be logged.

Detector and signal chain ​

Most compact vehicle-relevant LWIR cameras use uncooled microbolometers. A microbolometer pixel absorbs incoming IR energy, changes temperature slightly, and changes electrical resistance. The readout chain converts that response to digital counts:

text
scene radiance
-> lens/window transmission
-> microbolometer membrane heating
-> resistance/readout change
-> raw digital number
-> NUC/bad-pixel/temperature compensation
-> radiometric frame or display image
-> perception, tracking, mapping, or health monitor

Common signal products:

  • raw counts or 14/16-bit radiometric frames
  • non-uniformity-corrected frames
  • automatic-gain-control display imagery
  • apparent temperature image
  • dead-pixel mask and NUC/FFC event metadata

Automatic gain control and 8-bit display images can help a human view the scene, but they may destroy temporal and radiometric consistency. Direct thermal-inertial odometry, trend monitoring, and validation should preserve raw or radiometric data when the hardware exposes it.

NUC, FFC, and calibration tables ​

Thermal detectors have pixel-to-pixel gain and offset variation. Non-uniformity correction (NUC) or flat-field correction (FFC) compensates for this variation and for drift as the camera temperature changes.

A simplified pixel correction is:

text
DN_corrected(u) =
  gain(u, T_camera, table_id) * DN_raw(u)
  + offset(u, T_camera, table_id)

FLIR's OEM support material describes factory calibration against blackbody sources, pixel gain/offset terms, NUC tables, and table selection by camera internal temperature. FLIR's public NUC explainer also notes that NUC is used to adjust for detector drift as scene and environment change.

Operational implications:

  • A NUC/FFC event can create a frame jump, short blind interval, or feature discontinuity.
  • Warm-up state matters; a camera at the same ambient temperature can still have different internal thermal distributions.
  • Shutterless systems avoid a mechanical interruption but still need drift compensation and evidence that the compensation remains stable.
  • Bad-pixel maps and NUC table IDs are part of the measurement provenance, not only factory metadata.

Recent work on uncooled shutterless microbolometer cameras shows why long-term radiometric stability depends on camera thermal state, housing/optics contributions, and calibration model validity. Vehicle deployments should therefore monitor camera temperature and residual behavior rather than assuming factory calibration alone covers every environment.

Error budget and likelihood ​

A thermal measurement residual is usually not dominated by a single noise source. A useful error budget separates:

Error sourceAutonomy effect
NETD and temporal noiseSmall temperature differences become low-SNR features.
Fixed-pattern residualStripes, columns, or pixel texture become false features.
Emissivity errorApparent temperature differs from true surface temperature.
Reflected apparent temperatureShiny metal, glass, water, and wet pavement create false hot or cold objects.
Atmospheric pathFog, humidity, smoke, rain, steam, dust, distance, and hot exhaust change received radiance.
Optics/window emissionLens, germanium window, housing, heater, or contamination adds scene-independent bias.
NUC/FFC eventFrame discontinuity or temporary blindness breaks tracking.
Geometric calibration and timingThermal boxes, LiDAR points, and RGB features fail to align.
Dynamic range and saturationFires, brakes, sun-heated surfaces, or engines clip or compress nearby contrast.

For a detector, track, or SLAM residual, the likelihood should be conditioned on operating context:

text
R_thermal =
  R_detector_noise
  + R_radiometric_parameters
  + R_calibration
  + R_timing
  + R_environment
  + R_preprocessing

If the frame is contrast-only rather than radiometric, do not report a temperature covariance. Use detection confidence, feature residuals, or a context-specific measurement covariance instead.

AV relevance and fusion handoff ​

Thermal evidence is strongest when it complements another modality:

  • RGB provides texture, color, signs, and daytime semantics.
  • LiDAR provides geometry and range.
  • Radar provides velocity and weather-tolerant returns.
  • IMU and wheel odometry provide short-term propagation.
  • Maps provide static structure and exclusion zones.

Useful handoff patterns:

  • thermal detection plus LiDAR range to create a 3D personnel or vehicle track
  • thermal contrast plus radar/LiDAR evidence for fog, smoke, and night hazards
  • thermal health flags that downgrade night-speed ODD or trigger cleaning/maintenance
  • thermal-inertial odometry residuals with NUC-aware feature gating
  • thermal map layers tagged by season, time of day, weather, and equipment state

Thermal should not usually remove a good geometric detection by itself. It should add evidence, downgrade confidence when it is unhealthy, and provide unique cues such as heat-source state when the radiometric contract is valid.

Domain fit ​

DomainFitNotes
Road AVStrong night complementUseful for pedestrians, animals, vehicles, and glare/night fallback. Must handle sun-heated asphalt, glass, wet roads, and automotive cleaning/heating.
AirsideStrong safety complementUseful for personnel, engines, brakes, tires, APU/exhaust state, fuel/fire cues, and night operation. Must not confuse heat signature with geometry or certified separation.
Warehouse and logistics yardStrong in mixed lightingHelps around dock doors, cold-chain zones, forklifts, people, batteries, and charging. Reflections from metal racks and dock plates need validation.
Port, mining, construction, agricultureSituational to strongValuable for people, engines, fires, and machinery state; dust, mud, rain, exhaust, sun load, and lens contamination dominate health monitoring.
Delivery robot and campusSituationalHelps at night and near pedestrians, but low sensor height, weather, privacy constraints, and cost can outweigh benefit.
Indoor robot and tunnelStrong for smoke/darknessCan support search, inspection, and GNSS-denied odometry, but thermally uniform corridors and reflective surfaces are hard.

The cross-domain rule is to treat thermal as a radiometric or contrast sensor with a health contract, not as an airside-only jet-blast sensor or a universal night-vision replacement.

Failure modes and diagnostics ​

Failure modeSymptomDiagnostic
Wrong emissivityApparent temperature wrong while image looks plausible.Log emissivity assumption; test materials with reference targets or ASTM-style compensation.
Reflected heat sourceFalse hot/cold object on metal, glass, or wet surface.Multi-view consistency, polarization/geometry checks where available, and LiDAR/radar confirmation.
NUC/FFC jumpTrack or SLAM residual spike at correction event.Log NUC timestamps, table ID, camera temperature, and residual before/after event.
Thermal driftDetection thresholds or temperature estimates shift over time.Monitor camera body/FPA temperature, reference target residuals, and fleet drift trends.
AGC destroys consistencySame object changes intensity after scene histogram changes.Preserve raw/radiometric stream for algorithms; record display AGC as visualization only.
Low thermal contrastPerson, animal, or obstacle blends into background.Validate by ambient temperature, clothing, range, sun exposure, and weather.
Lens/window contaminationBlurred frame, vignetting, false cooling/heating pattern.Track sharpness, flat-field residual, cleaning events, and window heater state.
Geometric misregistrationThermal and LiDAR/RGB detections disagree.Check thermal intrinsics/extrinsics, target design, timestamp offset, and calibration drift.
Overclaimed temperature accuracySafety or maintenance decision uses unsupported temperature value.Separate radiometric mode, parameter completeness, and reference-check evidence from contrast-only perception.

Implementation checklist ​

  • Store raw or radiometric frames when possible; avoid using only 8-bit display images for validation.
  • Log radiometric parameters: emissivity, reflected apparent temperature, distance, humidity, atmospheric temperature, lens/window compensation, and calibration version.
  • Log NUC/FFC events, table IDs, bad-pixel maps, internal temperature, and warm-up state.
  • Keep camera intrinsics, extrinsics, timestamp semantics, and PTP/trigger provenance alongside every thermal stream used for fusion.
  • Validate by range, target size, material, emissivity, background temperature, weather, contamination, sun load, time of day, and domain.
  • Use reference targets or blackbody checks for radiometric claims.
  • Treat thermal maps as context-tagged layers, not permanent geometry.
  • Separate contrast-based perception evidence from temperature-measurement evidence in safety cases.

Sources ​

Public research notes collected from public sources.

Copyright © 2026 kvynlim