Thermal and Infrared Cameras for Airside Autonomous Vehicles

The Missing Sensor: Night Operations, Personnel Safety, and Jet Blast Detection

Last updated: 2026-04-11

Related docs: Thermal IR Radiometry First Principles, Event and Thermal Camera Models, and Night Operations Thermal Fusion.

---

Table of Contents

1. Why Thermal for Airside
2. IR Wavelength Bands
3. Detector Technologies
4. Automotive-Grade Products
5. Airside-Specific Applications
6. Thermal + Visible Fusion
7. Perception Models for Thermal
8. Deployment on NVIDIA Orin
9. Cost Analysis
10. Recommended Configuration
11. References

---

1. Why Thermal for Airside

1.1 The Night Operations Gap

Airport airside operations run 24/7. During night shifts, the critical perception challenges are:

Challenge	Visible Camera	LiDAR	4D Radar	Thermal Camera
Personnel at 50m (night)	Fails (poor lighting)	Detects (3D point)	Detects (velocity)	Excellent (body heat)
Personnel behind GSE	Fails (occlusion)	Partial	Detects (through gaps)	Detects (heat signature through gaps)
Hi-vis vest effectiveness	Day: good, Night: 84-88% AEB failure	N/A	N/A	N/A — detects person, not vest
Jet engine running vs stopped	Cannot distinguish	Cannot distinguish	Detects blade rotation	Clear (exhaust plume visible)
Fuel spill on apron	Difficult (transparent)	Cannot detect	Cannot detect	Detects (evaporative cooling)
De-icing fluid coverage	Cannot assess	Cannot assess	Cannot assess	Clear (thermal contrast)
Fire detection	Visible flame only	Cannot detect	Cannot detect	Detects (early, even small)
Tire/brake heat	Cannot see	Cannot see	Cannot see	Clear (overheating visible)

1.2 The Hi-Vis Paradox

From 70-operations-domains/airside/safety/ground-crew-pedestrian-safety.md:

Hi-vis clothing causes 84-88% failure rate in camera-based Automatic Emergency Braking at night due to retroreflective overexposure creating blooming artifacts.

Thermal cameras solve this completely. They detect the human body's thermal signature (~37°C) against the ambient background, regardless of clothing. Hi-vis vest is invisible to thermal — the person is always detectable.

1.3 Unique Airside Thermal Signatures

Object	Typical Temperature	Thermal Contrast	Detection Ease
Running jet engine exhaust	400-600°C	Extreme	Trivial
Jet exhaust plume (100m)	50-150°C above ambient	Very high	Easy
Human body	37°C (always)	10-25°C above ambient	Very easy
Recently-run APU	80-150°C	High	Easy
Hot brakes (after landing)	150-300°C	Very high	Easy
Fuel spill (evaporative cooling)	5-15°C below ambient	Medium	Moderate
De-icing fluid on surface	3-8°C below ambient	Low-medium	Moderate
Tire marks (friction heat)	5-20°C above ambient	Low-medium	Moderate
Vehicle engine compartment	80-110°C	High	Easy
Heated building (winter)	10-30°C above ambient	Medium	Easy
Metal GSE in sun	Up to 70°C above ambient	High	Easy (can cause false positives)

---

2. IR Wavelength Bands

2.1 Spectrum Overview

Visible  |  NIR   |  SWIR   |  MWIR   |  LWIR   |  VLWIR
0.4-0.7  | 0.7-1  | 1-2.5   | 3-5     | 8-14    | >14 μm
         |        |         |         |         |
 Cameras | Night  | See     | Cooled  | Uncooled|
         | vision | through | InSb    | Micro-  |
         | (Si)   | fog     | FPA     | bolom.  |
         |        | (InGaAs)|         | (VOx)   |

2.2 Band Comparison for Airside

Band	Wavelength	Detector	Cooling	Cost	Best For (Airside)
NIR	0.7-1.0 μm	Silicon CCD/CMOS	None	$100-500	Cheap night vision with active IR illumination
SWIR	1.0-2.5 μm	InGaAs	None/TEC	$5-20K	Seeing through fog, smoke, de-icing mist
MWIR	3-5 μm	InSb, MCT, T2SL	Cryo (-200°C)	$20-100K	Highest sensitivity, gas detection, long range
LWIR	8-14 μm	VOx microbolometer	None (uncooled)	$500-5K	Best overall for AV: personnel, engines, low cost

Recommended for airside AV: LWIR (8-14 μm)

• Uncooled = no cryocooler = lower cost, smaller, more reliable
• Peak thermal emission from room-temperature objects (Wien's law: 37°C body peaks at ~9.3 μm)
• Through atmosphere window (minimal absorption)
• Automotive-grade products available from multiple vendors

2.3 Atmospheric Windows

IR radiation must pass through atmosphere to reach camera.
Two main transmission windows:

Window 1: 3-5 μm (MWIR)
  - Good transmission
  - Hot objects (engines, exhaust) emit strongly here
  - BUT: requires cryogenic cooling

Window 2: 8-14 μm (LWIR)
  - Excellent transmission
  - Room-temperature objects emit strongly here
  - Uncooled detectors available
  - BEST for personnel detection

Between windows (5-8 μm): strong atmospheric absorption (water vapor)
  → Useless for outdoor imaging

---

3. Detector Technologies

3.1 Microbolometer (LWIR)

The dominant technology for automotive/AV thermal cameras:

Operating principle:
  IR photons → absorbed by thin film (VOx or a-Si)
  → temperature rise → resistance change → voltage change
  → no cooling needed (operates at room temperature)

Key specs (state-of-art 2025):
  Pixel pitch: 12 μm (high-end), 17 μm (standard), 25 μm (legacy)
  Resolution: 640×480 (VGA), 1024×768 (XGA), 1280×1024 (SXGA)
  NETD: 30-50 mK (can detect 0.03°C differences)
  Frame rate: 30-60 Hz (some up to 120 Hz for fast objects)
  Response time: 8-12 ms
  
Manufacturers (detector):
  - ULIS / Lynred (France) — market leader, VOx, 12μm pixel
  - DRS / Leonardo DRS (US) — military + automotive
  - BAE Systems (UK) — VOx
  - Raytron (China) — fast-growing, competitive pricing

3.2 Cooled InSb/MCT (MWIR)

For maximum performance (military, high-end surveillance):

Operating principle:
  IR photons → directly excite electrons in semiconductor
  → photocurrent proportional to IR intensity
  → requires cryogenic cooling to ~77K (-196°C)

Key specs:
  Pixel pitch: 10-30 μm
  Resolution: up to 2048×2048
  NETD: 10-20 mK (2-5x better than microbolometer)
  Frame rate: up to 1000 Hz
  
Cost: $20K-100K per camera (cooler dominates cost)
Size/weight: 2-5x larger than uncooled
MTBF: 5,000-10,000 hours (cooler limited)

NOT recommended for airside AV due to cost and reliability.
Exception: airport-wide surveillance from tower (shared infrastructure cost).

---

4. Automotive-Grade Products

4.1 FLIR / Teledyne FLIR

Teledyne FLIR is the dominant thermal camera brand, with dedicated automotive products:

Product	Type	Resolution	Pixel Pitch	NETD	Interface	Price (Est.)
Boson 640	Core module	640×512	12 μm	<40 mK	CMOS parallel, USB	$3-6K
Boson 320	Core module	320×256	12 μm	<50 mK	CMOS parallel, USB	$1.5-3K
Lepton 3.5	Micro module	160×120	12 μm	<50 mK	SPI, I2C	$200-400
ADK (Auto Dev Kit)	Full camera	640×512	12 μm	<40 mK	GigE, CAN	$5-10K
PathFindIR III	Automotive	320×240	25 μm	<60 mK	NTSC analog	$3-5K

Boson 640 is the recommended core for airside AV:

• Automotive-grade (-40°C to +80°C operating)
• 12 μm pixel pitch (state of the art for uncooled)
• 640×512 resolution — sufficient for pedestrian detection at 100m+
• Multiple lens options (4.9mm to 50mm focal length)
• Radiometric output (actual temperature values, not just relative)
• ITAR-controlled (US export restrictions may apply)

4.2 Seek Thermal

Product	Resolution	Interface	Price	Notes
Seek Starter Kit	206×156	USB-C	$150-250	Dev evaluation only
Seek Mosaic Core	320×240	MIPI CSI-2	$500-1K	Embeddable, automotive-intent
Seek Mosaic Pro	640×480	MIPI CSI-2, GigE	$2-4K	AV-grade, wide dynamic range

Seek Mosaic Pro is a strong alternative to FLIR Boson:

• Non-ITAR (no US export restrictions)
• MIPI CSI-2 interface (direct to Orin)
• Automotive-grade temperature range
• Competitive pricing

4.3 InfiRay (China)

Product	Resolution	NETD	Interface	Price	Notes
AT61	640×512	<30 mK	GigE	$2-4K	Best NETD in class
AT31	384×288	<35 mK	GigE, USB3	$1.5-3K	Mid-range
Micro III	256×192	<40 mK	USB-C, MIPI	$300-600	Compact module

InfiRay AT61 offers the best noise performance (<30 mK NETD) at competitive pricing. Subject to CFIUS/entity list considerations for US/EU military-adjacent applications.

4.4 Lynred (France, European Alternative)

Lynred (formerly Sofradir + ULIS) manufactures the detectors used by many thermal camera brands:

Product	Resolution	Pitch	Technology	Notes
PICO1024 Gen2	1024×768	12 μm	a-Si	Highest resolution uncooled
PICO640 Gen2	640×480	12 μm	a-Si	Standard automotive
Micro80 Gen2	80×80	12 μm	a-Si	Ultra-compact for array configs

Lynred detectors are used in cameras from Teledyne FLIR (Boson family), Opgal, Guide Sensmart, and others. For a European supply chain, Lynred + European integrator avoids ITAR/Chinese supply chain concerns.

---

5. Airside-Specific Applications

5.1 Personnel Detection at Night

The primary safety use case. 27,000 ramp accidents per year, many at night.

Thermal personnel detection at night:

Human body: ~33-35°C skin, ~37°C core
Ambient (night): -10°C to 25°C (seasonal)
ΔT: 8-47°C → ALWAYS detectable in LWIR

Detection range (640×512, 19mm lens):
  - Person detection: >200m
  - Person recognition: >100m
  - Identification: >50m

Detection range (640×512, 6.3mm lens, wide angle):
  - Person detection: >80m
  - Person recognition: >40m

Comparison with visible camera at night:
  Visible + apron lighting: detect person at ~20-30m (unreliable)
  Thermal LWIR: detect person at ~80-200m (highly reliable)
  Improvement: 3-8x detection range at night

5.2 Jet Blast Zone Detection

Jet engine exhaust temperature profile:

Engine type    | Exhaust temp (nozzle) | At 50m behind  | At 100m behind
─────────────  | ───────────────────── | ────────────── | ───────────────
Turbofan (CFM56)| 600°C               | ~80°C          | ~40°C
Turbofan (LEAP) | 550°C               | ~70°C          | ~35°C
APU             | 300-400°C           | ~40°C          | ~25°C
Turboprop       | 500°C               | ~60°C          | ~30°C

LWIR thermal camera can:
  1. Detect engine exhaust plume shape and extent
  2. Determine if engines are running or shutdown
  3. Estimate blast danger zone in real-time
  4. Detect unexpected engine start during pushback
  
This is IMPOSSIBLE with any other sensor:
  - Visible camera: cannot see hot air (transparent)
  - LiDAR: cannot detect gas (no solid particles)
  - Radar: partially detects turbulence, unreliable
  - Thermal: CLEARLY sees entire exhaust plume

5.3 De-Icing Fluid Monitoring

De-icing fluid (Type I/IV glycol):
  - Applied at 60-80°C
  - Cools to ambient over 5-30 minutes
  - Thermal camera shows:
    - Coverage completeness (is entire wing surface treated?)
    - Holdover time estimation (how fast is it cooling?)
    - Re-freeze detection (sudden temperature drop)
    
Application for autonomous de-icing GSE:
  - Thermal-guided spray pattern optimization
  - Automated coverage verification
  - Holdover time monitoring → signal departure readiness

5.4 FOD Detection

Foreign Object Debris detection via thermal contrast:

Day: FOD on concrete apron has different thermal mass
  - Metal FOD heats faster than concrete in sun
  - Rubber FOD absorbs more IR
  - Detection: moderate (depends on ΔT)

Night: FOD retains different heat than concrete
  - Metal cools faster → colder than concrete
  - Rubber retains heat → warmer than concrete
  - Detection: moderate-good

Best approach: thermal + visible fusion
  - Visible: detects color/shape contrast
  - Thermal: detects temperature contrast
  - Fusion: catches what either alone misses

5.5 Fire and Fuel Spill Detection

Fire detection:
  - Thermal camera detects flame at 600-1500°C
  - Also detects pre-fire conditions (overheating components)
  - Sub-second detection (vs minutes for smoke detectors)
  - Range: 100m+ for small fire

Fuel spill:
  - Jet fuel (Jet A-1) evaporates at ~40°C
  - Evaporative cooling creates thermal signature
  - Spill appears as cool region on warm concrete
  - Detection possible even for thin films
  
Combined with 4D radar:
  - Radar detects liquid surface reflectivity change
  - Thermal confirms evaporative cooling signature
  - Dual confirmation reduces false positives

---

6. Thermal + Visible Fusion

6.1 Fusion Approaches

Approach 1: Pixel-level fusion (early fusion)
  Thermal image + Visible image → Aligned → Combined multi-channel image
  → Single detection model on fused input
  
  Pros: Maximum information available to model
  Cons: Requires precise alignment (registration), different resolutions

Approach 2: Feature-level fusion (intermediate fusion)
  Thermal → Backbone → Features ──┐
                                   ├── Fusion module → Detection head
  Visible → Backbone → Features ──┘
  
  Pros: Handles misalignment better, can use different backbones
  Cons: More complex architecture

Approach 3: Decision-level fusion (late fusion)
  Thermal → Full detector → Detections ──┐
                                          ├── NMS/merge → Final detections
  Visible → Full detector → Detections ──┘
  
  Pros: Simplest, independent failure modes
  Cons: Cannot recover what either detector alone missed

6.2 Registration / Alignment

Thermal and visible cameras have different optics, resolution, and field of view. Alignment options:

Method	Accuracy	Effort	Notes
Stereo calibration	Sub-pixel	One-time setup	Use checkerboard heated with IR lamp
Deep homography	~2 pixel	Learned	Neural network predicts alignment
Feature matching	Variable	Automatic	Fails when thermal/visible features differ greatly
Co-located sensor	Perfect (shared optics)	None	Expensive dual-sensor cameras available

Recommended for airside: Stereo calibration (one-time) + deep homography refinement (handles small shifts from vibration/thermal expansion).

6.3 Key Datasets for Thermal Perception

Dataset	Year	Modalities	Annotations	Size	Focus
KAIST	2015	Visible + LWIR	Pedestrian bbox	95K pairs	Pedestrian detection, day/night
LLVIP	2021	Visible + LWIR	Pedestrian bbox	16.8K pairs	Low-light visible + infrared
FLIR ADAS v2	2022	Visible + LWIR	4 classes bbox	26.4K thermal	Automotive (vehicles, people, bikes, dogs)
M3FD	2022	Visible + LWIR	6 classes bbox	4.2K pairs	Multi-modal fusion, fog/night
DroneVehicle	2022	Visible + LWIR	Oriented bbox	56.9K pairs	Aerial vehicle detection
InfiRay MOOD	2023	Visible + LWIR	6 classes bbox	1.5K pairs	Automotive, Chinese driving

No airside thermal dataset exists. This is another gap and opportunity.

6.4 Night Performance Comparison

From KAIST and LLVIP benchmarks:

Method	Day (Visible)	Night (Visible)	Night (Thermal)	Night (Fusion)
YOLOv5 (pedestrian)	91.2%	62.8%	84.3%	89.7%
Faster R-CNN	89.1%	58.4%	81.2%	87.3%
Thermal-only YOLO	—	—	85.1%	—
Halfway Fusion	—	—	—	91.4%

Key insight: Thermal alone at night (84-85%) nearly matches visible during day (89-91%). Thermal+visible fusion at night (89-91%) achieves near-daytime performance.

---

7. Perception Models for Thermal

7.1 Pre-trained Models Available

Model	Framework	Thermal Dataset	Performance	License
YOLOv8 (fine-tuned on FLIR)	Ultralytics	FLIR ADAS v2	~80% mAP pedestrian	AGPL/Enterprise
YOLOX-Thermal	YOLOX	KAIST	~76% mAP pedestrian	Apache 2.0
Thermal-DETR	DETR variant	LLVIP	~82% mAP pedestrian	Research
ThermalDet	Custom	Multi-dataset	~85% mAP pedestrian	Research

7.2 Domain Adaptation for Airside

Road thermal datasets (KAIST, FLIR ADAS) have different characteristics from airside:

Aspect	Road Thermal	Airside Thermal
Pedestrians	Casual clothing, varied poses	Hi-vis + PPE, airport-specific gear
Vehicles	Cars, trucks, buses	GSE (tractors, belt loaders, fuel trucks)
Background	Asphalt, buildings, trees	Concrete, aircraft, jet bridges
Temperature	Wide ambient range	Similar, plus jet exhaust heat
Viewing angle	Dash-mounted (1.2-1.5m)	Various (1-3m on GSE)

Transfer learning strategy:

1. Start with YOLO/DETR pre-trained on COCO (visible)
2. Fine-tune on FLIR ADAS thermal dataset (general thermal)
3. Fine-tune on small airside thermal dataset (100-500 annotated frames)
4. Expected: >80% mAP for personnel detection with just 200 airside frames

7.3 Thermal Object Detection on Orin

Pipeline: Thermal camera → GStreamer → TensorRT inference → ROS topic

GStreamer pipeline for thermal camera:
  gst-launch-1.0 \
    v4l2src device=/dev/video_thermal ! \
    video/x-raw,format=GRAY16_LE,width=640,height=512,framerate=30/1 ! \
    videoconvert ! \
    appsink name=thermal_sink

TensorRT model:
  Input: 640×512 single-channel (16-bit thermal)
  Model: YOLOv8s fine-tuned for thermal
  Output: bounding boxes + class labels
  
  Latency on Orin (FP16): ~8-12ms per frame
  Memory: ~500 MB
  Power: ~5W

ROS integration:
  /thermal_raw (sensor_msgs/Image, encoding=mono16)
  /thermal_detections (vision_msgs/Detection2DArray)
  /thermal_overlay (sensor_msgs/Image — visualization with boxes)

---

8. Deployment on NVIDIA Orin

8.1 Camera Interface Options

Interface	Bandwidth	Latency	Orin Support	Recommended For
MIPI CSI-2	Up to 2.5 Gbps/lane	<1ms	Native (16 CSI-2 ports)	Direct connection, lowest latency
GigE Vision	1 Gbps	1-5ms	Via NIC (PCIe)	Standard industrial, longer cable
USB3	5 Gbps	1-3ms	Native USB3.2	Development, less reliable for production
Analog (NTSC)	N/A	~33ms	Via capture card	Legacy cameras only

Recommended: MIPI CSI-2 for Orin integration (direct, lowest latency, no additional hardware). Boson 640 and Seek Mosaic both support MIPI CSI-2.

8.2 Thermal Camera Integration Diagram

Orin AGX Developer Kit
  ├── CSI Port 0-3: Visible cameras (4× surround)
  ├── CSI Port 4-5: Thermal cameras (2× front + rear)
  │     ├── FLIR Boson 640 (front, 50mm lens) — long range
  │     └── Seek Mosaic Pro (rear, wide angle) — close range
  ├── Ethernet: LiDAR data (4-8× RoboSense)
  ├── CAN: Vehicle bus
  └── USB: IMU, GPS

Processing pipeline:
  CSI thermal → ISP bypass (raw 14-bit) → NvBufSurface 
  → TensorRT inference (YOLO thermal, 8ms) → ROS topic
  
  Total thermal pipeline latency: <15ms
  Additional GPU load: ~3-5W, ~500 MB

8.3 Multi-Spectral BEV Fusion

For full-stack integration, thermal can be fused into the BEV perception pipeline:

Visible cameras (6-8) → Image backbone → BEV features
Thermal cameras (2-4) → Thermal backbone → BEV features (separate)
LiDAR (4-8) → PointPillars → BEV features

All BEV features → Multi-modal BEV fusion (BEVFusion-style)
                → Unified 3D perception

Thermal channels: Add thermal BEV as additional modality
  - Same LSS (Lift-Splat-Shoot) projection as visible cameras
  - But thermal provides complementary information:
    - People detected even when visible fails (night, hi-vis blooming)
    - Engine status visible (running/stopped)
    - Hot exhaust zones mapped

---

9. Cost Analysis

9.1 Per-Vehicle Cost

Component	Quantity	Unit Cost	Total	Notes
FLIR Boson 640 (core)	2	$4-6K	$8-12K	Front + rear
Lens (19mm, f/1.0)	2	$500-1K	$1-2K	Matched to Boson
Housing (IP67, automotive)	2	$500-1K	$1-2K	Weatherproof enclosure
MIPI CSI-2 adapter board	2	$200-400	$400-800	Interface to Orin
Calibration + integration	1	$2-5K	$2-5K	Engineering time
Total per vehicle			$12.4-21.8K

Alternative (budget):

Component	Quantity	Unit Cost	Total
Seek Mosaic Pro (core)	2	$2-4K	$4-8K
Lens + housing	2	$1-1.5K	$2-3K
Integration	1	$2-3K	$2-3K
Total per vehicle			$8-14K

9.2 Infrastructure Thermal Cameras

For fixed infrastructure (stand monitoring, taxiway surveillance):

Use Case	Camera	Quantity/Airport	Cost Each	Total
Stand personnel monitoring	FLIR A700 (fixed)	20-40 (1 per stand)	$5-8K	$100-320K
Taxiway intersection	PTZ thermal	5-10	$8-15K	$40-150K
Wide-area surveillance	Pan-tilt thermal	2-4 (on towers)	$15-30K	$30-120K
Total infrastructure				$170-590K

9.3 ROI for Night Safety

Metric	Without Thermal	With Thermal
Night pedestrian detection range	20-30m (visible)	80-200m (thermal)
Night detection reliability	60-70%	95%+
Ramp incident reduction	Baseline	Est. 30-50% fewer incidents
Annual ramp incident cost (per airport)	$2-10M	$1-5M
Annual savings		$1-5M
Payback period (per vehicle)		<1 year

---

10. Recommended Configuration

10.1 Vehicle-Mounted Thermal (2 cameras per GSE)

Position 1: Front-facing, long range
  Camera: FLIR Boson 640 + 19mm f/1.0 lens
  FoV: 32° × 26°
  Range: 200m+ person detection
  Purpose: Forward path safety, approaching aircraft, jet blast

Position 2: Rear/side, wide angle  
  Camera: FLIR Boson 640 + 6.3mm f/1.0 lens
  FoV: 95° × 75°
  Range: 50m+ person detection
  Purpose: Close-range personnel, reversing safety, side clearance

Optional Position 3-4: Additional side-facing for 360° coverage
  Camera: Seek Mosaic Pro (lower cost)
  Purpose: Complete surround thermal coverage

10.2 Integration Priority

Priority	Integration Step	Effort	Benefit
1	Mount 2× thermal on one test vehicle	2-4 weeks	Night personnel detection proof-of-concept
2	Train YOLO on FLIR ADAS + small airside set	1-2 weeks	Thermal person/vehicle detector
3	Late fusion with existing LiDAR pipeline	2-4 weeks	Combined day/night detection
4	BEV fusion (thermal + LiDAR + visible)	2-3 months	Full multi-spectral perception
5	Jet blast zone estimation from thermal	1-2 months	Unique safety feature

10.3 Sensor Suite Update

Adding thermal to the recommended airside sensor suite (from 90-synthesis/master/master-synthesis.md):

Sensor	Qty	Purpose	Airside Rationale
Cameras (surround)	6-8	Primary perception	Rich semantic info
Thermal LWIR	2-4	Night safety, jet blast	Personnel at night, engine status
LiDAR (128-ch)	1	3D geometry	Large objects, localization
4D Radar	2-4	Velocity, all-weather	Rain, de-icing, fog
RTK-GNSS	1	Localization	cm-level in open areas
ADS-B receiver	1	Aircraft awareness	Real-time aircraft positions
IMU	1	Dead reckoning	Bridge GPS gaps

---

11. References

Products

• FLIR Boson: https://www.flir.com/products/boson/
• FLIR ADK: https://www.flir.com/products/adk/
• Seek Thermal Mosaic: https://www.thermal.com/mosaic-core.html
• InfiRay AT Series: https://www.infiray.com/thermal-modules.html
• Lynred PICO series: https://www.lynred.com/products

Datasets

• KAIST Multispectral Pedestrian: https://soonminhwang.github.io/rgbt-ped-detection/
• LLVIP: https://bupt-ai-cz.github.io/LLVIP/
• FLIR ADAS v2: https://www.flir.com/oem/adas/adas-dataset-agree/
• M3FD: Multi-modal Fusion Dataset for autonomous driving

Papers

• Hwang et al., "Multispectral Pedestrian Detection: Benchmark Dataset and Baseline," CVPR 2015
• Jia et al., "LLVIP: A Visible-Infrared Paired Dataset for Low-Light Vision," ICCV 2021
• Zhang et al., "Multispectral Fusion for Object Detection with Cyclic Fuse-and-Refine Blocks," ICIP 2020
• Cao et al., "Attention-Guided Multi-modal and Multi-scale Fusion for Multispectral Pedestrian Detection," IEEE TITS 2023

Related Documents

Topic	Document
4D radar (all-weather primary)	20-av-platform/sensors/4d-radar.md
Ground crew safety (hi-vis paradox)	70-operations-domains/airside/safety/ground-crew-pedestrian-safety.md
Adverse conditions (de-icing, fog)	60-safety-validation/verification-validation/robustness/airside-adverse-conditions.md
FOD and jet blast	70-operations-domains/airside/operations/fod-and-jetblast.md
Production perception systems	30-autonomy-stack/perception/overview/production-perception-systems.md
Sensor fusion architectures	30-autonomy-stack/perception/overview/sensor-fusion-architectures.md
NVIDIA Orin technical	20-av-platform/compute/nvidia-orin-technical.md
Infrastructure cooperative perception	30-autonomy-stack/perception/overview/infrastructure-cooperative-perception.md