Why a Thermal Camera Rated for Hundreds of Meters Might Alert You at Twenty

Most teams buy thermal on two numbers. Resolution and NETD go into the comparison spreadsheet, the lowest NETD wins, and the purchase order goes out.

Then the camera gets installed and the alerts do not arrive when anyone expected. Nothing is faulty. The spec sheet was accurate and the deployment still disappointed, because the numbers on it were answering a different question from the one you were asking.

Thermal is worth understanding properly, because as a sensor it does things no visible camera can. It just has a set of assumptions baked into every published figure, and those assumptions are rarely the conditions on your site.

Key Takeaways

• Thermal imaging for security uses long-wave infrared, roughly 8 to 14 micrometers, and reads emitted heat rather than reflected light.
• The headline NETD figure is usually measured after digital noise reduction. One manufacturer's own white paper shows roughly 60 mK before processing and 20 mK after.
• DRI distances describe what a human operator can distinguish on screen. Published AI detection ranges are far shorter, sometimes by a factor of four or more.
• Thermal contrast, not distance, is the real constraint. When the target and background sit at similar temperatures, range collapses.
• Thermal cannot see through glass, and objects that do not emit heat may not be detected at all.
• Radiometric models are a different product class, and several now publish temperature data over MQTT.

What the Sensor Is Actually Measuring

The infrared band runs from about 0.75 to 1,000 micrometers. Near and short-wave infrared work with reflected light, which is why those images look broadly like visible ones.

Mid and long-wave infrared are different. They read radiation emitted by the object itself, which is why a person shows up in total darkness with no illumination at all.

Security cameras almost always use long-wave infrared, typically 8 to 14 micrometers, captured by an uncooled microbolometer built on vanadium oxide or amorphous silicon. Manufacturing advances brought the cost down far enough for general commercial use.

The sensor specs worth knowing are pixel pitch, usually 12 or 17 micrometers, and resolution. QVGA at 320x240 and VGA at 640x480 remain common, with XGA and SXGA increasingly available.

For teams already running physical environment monitoring across data centers or production floors, this is a familiar category of problem. A thermal camera is a sensor that emits events into your stack, and like any sensor its datasheet performance was established in a lab.

NETD, and the Number You Are Actually Being Quoted

Noise equivalent temperature difference is the sensor's temperature resolution, expressed in millikelvin. A 20 mK camera can distinguish temperature differences smaller than 0.02 degrees Celsius, and lower is better.

Here is the part that changes how you read a comparison table. Hanwha Vision's own thermal white paper states that applying digital noise reduction cuts noise by roughly 65%, taking a sensor from around 60 mK before processing to around 20 mK after.

That is not a trick, and every manufacturer does something similar. It does mean the quoted figure describes the processed output under one set of conditions rather than the raw physics of the detector.

NETD also shifts with measurement temperature, lens F-number and integration time, all of which vendors specify for their own test setup. The same white paper is blunt that chasing a low NETD in isolation can mislead.

The metric that actually predicts whether an operator can resolve something is MRTD, minimum resolvable temperature difference, which folds in resolution, lens quality, processing and the observer. Low NETD is necessary for good MRTD but nowhere near sufficient.

There is a practical consequence for procurement. Two cameras quoting an identical 20 mK will not perform identically if one reaches that figure through a better detector and the other through more aggressive noise reduction, because heavy processing can smooth away the small thermal detail you needed.

So ask for the test conditions alongside the number. Reference temperature and lens F-number are the two that move the result most, and any vendor who measures properly will have them documented.

DRI Distance Is Not Your Alert Distance

Every quote you receive will lead with DRI, the detection, recognition and identification framework the US Army defined back in the 1950s. Detection means separating an object from its background, recognition means classifying it as human or vehicle, identification means picking out specifics.

Specifying a thermal security camera works best when you start from the analytics detection range you actually need and work backwards to lens and sensor, rather than starting from the headline DRI figure and hoping the software keeps up.

The gap between the two is larger than most buyers expect, and to its credit Hanwha Vision publishes both sets of numbers in the same document. On one of its 384-pixel models, the calculated DRI detection distance for a person is 115 meters. The measured AI object detection range for a person on that same model is 26 meters.

Look closely and a useful rule of thumb falls out. That 26 meter analytics figure sits much closer to the model's DRI recognition distance of 29 meters than to its detection distance.

So plan against recognition, not detection. If your analytics need to classify something as a person before raising an alert, the recognition column is the honest number to design around.

The Assumptions Buried in Every Range Figure

Thermal contrast is the real constraint, and it is the one nobody puts on a spec sheet. A clear thermal image requires a meaningful temperature difference between the target and its background, and when that difference narrows the effective range falls away.

This changes through the day and across seasons at a fixed site. A wall that has absorbed sun all afternoon will sit close to body temperature by early evening, and detection performance in that scene is simply not what it was at dawn.

Weather is more nuanced than the marketing suggests. Long-wave infrared genuinely outperforms visible in smoke, dust and haze, but when fog droplets reach 10 to 20 micrometers the attenuation is comparable to other bands. Snow and rain reduce infrared transmittance too.

Two hard limits are worth writing into any design document. Thermal cannot image through glass, because glass reflects infrared rather than passing it, so a camera behind a window sees the window.

And thermal detects heat, so objects that emit none may not register. Hanwha Vision's own documentation gives bicycles and parked cars as examples. Because thermal carries no color information, detection also degrades when objects overlap.

Radiometric Is a Different Product

If your interest is equipment condition rather than intruders, you want a radiometric model, and the distinction matters at procurement. Detection models flag changes within a region and trigger alarms. Radiometric models report actual temperature values.

The practical constraints are unfamiliar to most IT teams. Emissivity governs accuracy, and polished metal sits at 0.2 or below, which makes temperature readings on a clean metal surface close to meaningless. Oxidized or corroded metal reads 0.8 or higher.

Geometry matters too. Emissivity stays broadly stable between 0 and 45 degrees of viewing angle and degrades beyond that, so mount within 45 degrees of your target. For a reliable reading, the target should cover at least a 3x3 block of pixels rather than a single one.

The integration story is better than you might expect. Several models publish region-of-interest maximum, minimum and average temperatures over MQTT following the ONVIF event standard, which drops neatly into an existing observability pipeline rather than requiring a separate console.

A Commissioning Checklist That Survives Contact With the Site

Allow for warm-up. Sensors need to reach thermal equilibrium before readings and detection are trustworthy, and around 30 minutes is typical.

Expect brief interruptions. Single-point non-uniformity correction closes a shutter in front of the lens periodically to recalibrate, which momentarily freezes the image. Worth knowing before someone raises an incident about it.

Run a proof of concept in the actual environment, at the times of day you care about. The manufacturer recommends exactly this, which tells you how much published ranges depend on local conditions.

Finally, consider bi-spectrum. Pairing thermal and visible sensors in one housing lets thermal do the detecting while the visible channel supplies the color and detail needed to decide whether to act, with handover to a PTZ camera for a closer look.

The Short Version

Thermal earns its place where visible cameras fail: total darkness, smoke, dust and long open perimeters with no lighting budget. That is a real capability and nothing else does it as cheaply.

Just buy it on the right numbers. Ask for the analytics detection range rather than DRI, ask what conditions the NETD figure was measured under, and insist on a site trial before the invoice.

FAQ

Can a thermal camera identify a face or read a license plate? No. Thermal produces a heat map with no color or fine surface detail, so identification in the DRI sense means distinguishing a person from a dog, not knowing which person. Pair it with a visible camera if you need evidential detail.

Does thermal really see through fog? Partly. It performs well in smoke, dust and light haze, but when fog droplets reach roughly 10 to 20 micrometers, long-wave infrared attenuates much like other wavelengths.

Why can't I mount it behind a window? Glass reflects infrared rather than transmitting it, so the camera images the glass and any heat reflected in it. Thermal cameras need an unobstructed view or a purpose-built germanium window.

Is a lower NETD always better? It helps, but only alongside resolution, lens quality and processing. Confirm the test conditions behind any quoted figure, since values measured at different F-numbers or reference temperatures are not directly comparable.

How many pixels do I need on a target? For temperature measurement, plan for at least 3x3 pixels on the object. For detection and alerting, work from the vendor's published analytics range rather than calculating from pixel counts yourself.