THERMAL AI
TARGETING

Thermal imaging defeats visual camouflage and operates in zero visibility. Combined with YOLOv8 AI trained on thermal signatures, it provides automatic target detection 24 hours a day regardless of weather or light conditions.

How Thermal Cameras See

Everything emits infrared radiation based on its temperature. A human body at 37°C emits different wavelengths than cold ground at -10°C. A infrared camera (LWIR, Long-Wave Infrared) detects these differences and creates an image where hot objects appear bright and cold objects appear dark.

What this means operationally: a tank hidden under camouflage netting is invisible to a normal camera but glows like a lighthouse on infrared. Its engine (80-150°C), exhaust (200-400°C), and recently fired gun barrel (300-600°C) are dramatically hotter than everything around them. The heat-based camera sees through the camouflage because the netting is thin enough that heat radiates through it.

AI On Thermal

YOLOv8 trained on infrared images works identically to visual detection (see §6.1). Instead of recognizing color patterns, it recognizes heat patterns. The training process is the same: show the AI thousands of heat-based images with labeled targets, and it learns to automatically spot thermal signatures that match vehicles, personnel, and installations.

The recommended infrared camera for drone operations is the Infiray T2S+ (€250): 256×192 pixels, LWIR (8-14 μm), 19 μm pixel pitch, USB-C interface. It weighs 28g and connects directly to the Jetson Orin Nano. Lower resolution than visual cameras but sufficient for vehicle detection at 100-300m AGL.

Thermal Contrast Detection Range — Derivation

Thermal detection range is fundamentally limited by target-background temperature contrast divided by atmospheric attenuation. Starting from the radiometric range equation, we derive the maximum detection distance at which a target presents sufficient contrast above the camera's Noise Equivalent Temperature Difference (NETD).

R_max = √[(A_target · |T_target⁴ − T_bg⁴| · ε · σ) / (NETD · ρ)] · atm(λ, R)

Where:
    A_target    = target cross-sectional area (m²)
    T_target    = target apparent temperature (K)
    T_bg        = background temperature (K)
    ε           = emissivity (~0.95 for painted metal)
    σ           = Stefan-Boltzmann constant (5.67 × 10⁻⁸ W·m⁻²·K⁻⁴)
    NETD        = camera noise-equivalent temperature difference (~0.05 K for T2S+)
    ρ           = minimum signal-to-NETD ratio for detection (~5, rule of thumb)
    atm(λ, R)   = atmospheric transmission at wavelength λ over range R
                  (~0.85/km in LWIR at sea level, worse in humidity/rain)

Worked example — idle main battle tank at 100 m AGL. Engine compartment temperature 80°C (353 K) against forest background 15°C (288 K). Target cross-section visible from above: roughly 3 m × 7 m = 21 m² (T-72 dimensions). Using ε = 0.95, σ = 5.67e-8, NETD = 0.05 K, ρ = 5, clear weather atmospheric loss 0.85/km:

The geometric conclusion: from Fischer 26's typical 100-300 m AGL reconnaissance altitude, all vehicle-sized targets within the camera's field of view are well within radiometric detection range even in degraded atmospheric conditions. The binding constraint on thermal ISR from Fischer 26 is not range — it is pixel count per target, which is addressed by the tele-lens option on Fischer 26E.

Why This Matters Operationally

Thermal imaging matters because it detects what visual imaging cannot. Under concealment (camouflage netting, foliage cover, smoke, or night darkness), a running vehicle remains visible through its thermal signature — engine compartment, exhaust, recently-fired gun barrel. The visual camera sees nothing; the thermal camera sees a clearly-differentiated heat source. Against Russian doctrine that explicitly invests in visual camouflage (maskirovka), thermal becomes the primary sensor modality during ISR operations.

Combined with YOLOv8 running locally on the Jetson Orin Nano at 30 FPS, thermal ISR becomes automated: the operator does not need to scan every frame by eye. YOLOv8 flags potential targets with confidence scores, Dempster-Shafer fusion combines thermal with visual for confirmation, and only flagged detections require human review. A single Fischer 26 operator with thermal AI can effectively surveil an area that would require a platoon of visual-only observers in traditional doctrine.

External source: Värmekamera – Wikipedia

Open the interactive Threat Fusion Dashboard →

Sources

Physical laws and parameters. Infrared emission by temperature — Stefan-Boltzmann law (E = σT⁴). Spectral distribution — Planck's law. Atmospheric transmission windows (8–14 μm for LWIR) — standard atmospheric spectroscopy. Dempster-Shafer fusion formula for independent sensors — Dempster/Shafer evidence theory. Formal verification: thermal contrast detection ranges and YOLOv8 inference latency are verified in provable_claims.py (proofs YOLOV8_INFERENCE_LATENCY and DS_FUSION_THREE_SOURCES).

Parameter sources. Infiray T2S+ characteristics (256×192, 19 μm, 8–14 μm LWIR, 28 g) — manufacturer datasheet at infiray.com. Typical target temperatures (tank engine 60–100 °C, exhaust 200–400 °C, gun barrel 300–600 °C) — open research on infrared camouflage (FLIR Systems whitepaper on Infrared Signature Management). Jetson Orin Nano — NVIDIA specification. FLIR ADAS thermal datasets — public scientific base (14,000 annotated images).

Operational estimates — not validated by FSG-A in the field. Detection ranges (2 km tank, 1.5 km truck, 500 m person) are calculated from thermal contrast and Infiray optics, not measured by FSG-A on real targets. The "30–70% range reduction in fog/rain" is estimated from atmospheric absorption, not measured. All mAP figures are reproductions of published YOLOv8n benchmarks, not independently validated by FSG-A. False positives from sun-heated rocks at 40–60 °C are a descriptive scenario based on physics, not documented statistically.

External standards and references. Infiray T2S+ thermal module specifications (infiray.com). YOLOv8 thermal training guides (Ultralytics community, 2024). "Infrared Signature Management" — FLIR Systems whitepaper. FOI research on thermal detection from UAS (2024). FSG-A has no thermal flight data — temperatures are from public sources.

False Positive Mitigation

The most common thermal false positive in Swedish terrain: sun-heated dark rocks that reach 40-60°C on summer afternoons. YOLOv8 occasionally classifies large heated rocks as vehicles because the thermal signature shape and temperature overlap with cold-engine vehicles. Mitigation strategies: temporal filtering (a rock does not move — if the detection remains stationary for 5+ minutes without vehicle characteristics on visual camera, downgrade confidence by 30 percent), seasonal calibration (adjust detection thresholds higher in summer when background temperatures are elevated), and dual-camera verification (visual camera shows rock texture, not vehicle hull — the fusion algorithm catches what thermal alone misses).

Try the interactive Dempster-Shafer Fusion Calculator →

Implementation

# Dual-Camera Fusion — Visual + Thermal
# pip install numpy
# pip install opencv-python
import cv2, numpy as np

def fuse_visual_thermal(visual_dets, thermal_dets, max_dist_px=50):
    """Match visual and thermal detections of same target."""
    fused = []
    
    for v_det in visual_dets:
        best_match = None
        best_dist = max_dist_px
        
        for t_det in thermal_dets:
            # Compare bounding box centers
            dist = np.sqrt((v_det.cx - t_det.cx)**2 + (v_det.cy - t_det.cy)**2)
            if dist < best_dist:
                best_dist = dist
                best_match = t_det
        
        if best_match:
            # Fuse: visual gives vehicle model, thermal confirms running engine
            fused_conf = 1 - (1-v_det.conf) * (1-best_match.conf)  # Dempster-Shafer
            fused.append({
                "class": v_det.cls,          # Visual: "T-72"
                "thermal_active": True,       # Engine running
                "confidence": fused_conf,     # Combined: higher than either alone
                "position": v_det.position
            })
    
    return fused

The dual-sensor approach combines strengths of both imaging modalities. Visual cameras provide shape, color, and texture — enough to distinguish a T-72 from a BMP-2 at 500 meters. The infrared sensor provides heat signature — distinguishing a running engine from a cold abandoned vehicle. Neither alone gives complete information. Combined through Dempster-Shafer fusion, the two sensors produce confidence levels that exceed either individual sensor by significant margins.

The sensor fusion architecture processes both visual and infrared streams simultaneously on the Jetson Orin Nano, allocating GPU resources dynamically based on ambient light conditions. During daylight operations the visual stream receives 60 percent of processing capacity for higher-resolution classification. At dusk and dawn, allocation shifts toward equal processing. In full darkness, the infrared stream receives priority while the visual camera switches to near-infrared illumination mode for supplementary data that aids in distinguishing closely similar vehicle types.

The dual-sensor approach used by Fischer 26 and Lisa 26 combines two distinct imaging modalities for maximum detection reliability. The visual spectrum provides shape, color, and texture details that enable vehicle model identification. The infrared spectrum provides heat signature data that indicates operational status. Neither sensor alone provides complete information — the combination through Dempster-Shafer fusion produces confidence levels significantly exceeding either individual sensor.

Thermal Detection Performance (Published Benchmarks)

YOLOv8n is typically trained on the FLIR ADAS thermal dataset (14,000 annotated thermal images). Fine-tuning on 500 custom Nordic thermal images (winter conditions, cold vehicle signatures) is a recommended specialization path; FSG-A has not yet collected such a dataset. Figures below are published reference performance for YOLOv8n on thermal data, reproduced by operators who have run this pipeline — not FSG-A field measurements.

Thermal is COMPLEMENTARY to visual, not a replacement. Use both cameras simultaneously: visual for daylight classification, thermal for night detection and through-smoke capability. Lisa 26 fuses detections from both sensors — a target detected by both has higher confidence than either alone.

Related Chapters