A thermal camera is an imaging sensor that forms a picture from the infrared radiation emitted by objects according to their temperature, rather than from reflected visible light. In robotics, thermal cameras belong to the family of Sensors that extend perception beyond what RGB cameras can do: they detect people, animals, hot machinery and fire in complete darkness, through smoke and in glare.
Most robotic thermal cameras work in the long-wave infrared (LWIR) band of about 8 to 14 micrometers, where objects near room temperature emit most strongly. They appear on inspection robots, search-and-rescue platforms, agricultural drones, security patrols and autonomous vehicles, usually fused with visible, depth or LiDAR data to give a richer model of the scene.
How a thermal camera senses heat
Every object above absolute zero emits thermal radiation, and the amount rises steeply with temperature. A thermal camera focuses this radiation with a lens made of germanium, chalcogenide glass or silicon, because ordinary glass blocks long-wave infrared. The lens projects the scene onto a focal plane array of detector pixels.
In most robotic cameras the detector is an uncooled microbolometer: each pixel is a tiny membrane of vanadium oxide or amorphous silicon whose electrical resistance changes as it warms. The readout circuit converts these changes to a temperature image. Higher-end systems use cooled photon detectors such as indium antimonide or mercury cadmium telluride, which respond faster and more sensitively but need a cryogenic cooler.
The table below summarizes the main detector technologies and their trade-offs for mobile robots.
| Detector type | Typical band | Strengths | Limitations |
|---|---|---|---|
| Uncooled microbolometer | LWIR, 8 to 14 µm | Compact, low power, affordable | Slower response, lower sensitivity |
| Cooled InSb or MCT | MWIR, 3 to 5 µm | Very sensitive, fast frame rates | Costly, heavy, cooler needs maintenance |
| Thermopile array | LWIR | Very cheap, simple interface | Few pixels, coarse image |
Key specifications of a thermal camera
Choosing a thermal camera for a robot means balancing resolution, sensitivity, frame rate and radiometric capability against size, weight and power. Resolution is far lower than in visible cameras: common uncooled sensors offer 160 × 120, 320 × 256 or 640 × 512 pixels.
The following table lists the parameters that matter most and their typical values for uncooled units.
| Parameter | Typical value | What it affects |
|---|---|---|
| Resolution | 160 × 120 to 640 × 512 | Detection range and detail |
| NETD (thermal sensitivity) | About 30 to 60 mK | Ability to see small temperature differences |
| Frame rate | 9 Hz to 60 Hz | Motion blur, control loop latency |
| Field of view | About 25° to 90° | Coverage versus range |
| Radiometric accuracy | About ±2 °C or ±2 % | Trust in absolute temperature readings |
| Pixel pitch | 8 to 17 µm | Optics size and resolution per area |
Noise-equivalent temperature difference (NETD) is the smallest temperature contrast the sensor can resolve above its noise floor. A radiometric camera additionally reports calibrated temperature for each pixel, which inspection robots need; a non-radiometric camera only provides relative contrast, which is enough for detection tasks.
Thermal cameras compared with other robot vision sensors
A thermal camera complements rather than replaces other sensors. It excels at finding warm targets regardless of lighting, but it provides no direct range and little surface texture. Many robots therefore combine it with a depth camera or LiDAR.
| Sensor | What it measures | Works in darkness | Main weakness |
|---|---|---|---|
| Thermal camera | Emitted LWIR radiation | Yes | Low resolution, no direct depth |
| RGB camera | Reflected visible light | No | Fails in dark, smoke and glare |
| Depth camera | Distance via stereo or structured light | Partly, if active | Short range, sunlight interference |
| LiDAR | Distance via laser time of flight | Yes | Sparse, no temperature data |
| Infrared sensor | Single-point IR intensity | Yes | No image, only presence or proximity |
The distinction from a simple infrared sensor is important: a thermal camera produces a two-dimensional temperature image, while a passive infrared detector reports only change or presence. Near-infrared cameras with active illuminators are also different, because they still rely on reflected light, not emitted heat.
Typical uses of thermal cameras in robotics
Thermal imaging is chosen wherever heat itself carries the information or where visible light is unreliable. The table below maps common robotic applications to what the camera contributes.
| Application | Robot platform | Role of thermal imaging |
|---|---|---|
| Search and rescue | Drones, tracked ground robots | Find people in rubble, smoke or night |
| Industrial inspection | Legged or wheeled patrol robots | Detect hot bearings, cables and leaks |
| Firefighting | Remote-controlled vehicles | Locate hotspots through dense smoke |
| Agriculture | Field drones | Map crop water stress and livestock health |
| Autonomous driving | Cars and delivery robots | Detect pedestrians at night and in glare |
In sensor fusion pipelines, thermal images are registered to RGB or LiDAR frames through extrinsic calibration, often with heated checkerboard targets. Neural networks trained on thermal datasets then detect and segment people or vehicles, and the fused result feeds navigation or alarms.
Limitations and common pitfalls
The most common misconception is that a thermal camera reads true temperature everywhere. In reality the reading depends on emissivity, the efficiency with which a surface radiates. Polished metals have low emissivity and mostly reflect the surroundings, so a shiny hot pipe can look cold, and a reflection can look like a heat source.
- Glass and water block LWIR, so a thermal camera cannot see through windows and sees mostly the glass surface.
- Atmospheric absorption from humidity and rain reduces contrast and range.
- Thermal crossover at dawn and dusk can make targets and background the same temperature.
- Non-uniformity correction briefly freezes the image when the internal shutter closes, which robots must tolerate.
- Export controls apply to high-resolution or fast cameras in many countries, which affects product design and shipping.
Engineers should set correct emissivity and distance parameters for radiometric work, avoid relying on thermal data alone for obstacle avoidance, and plan for shutter interruptions in real-time control loops.
In simple terms
A thermal camera sees warmth instead of light. Everything gives off a little invisible heat glow, and warmer things glow more. The camera turns that glow into a picture where people, engines and hot wires stand out clearly, even at night or in smoke.
For a robot, this is like gaining a second pair of eyes that never needs a lamp. It cannot read signs or see through a window, but it can spot a person or an overheating machine that an ordinary camera would miss.
A practical example
A chemical plant deploys a four-legged inspection robot that walks a fixed route through its pump hall every night. The robot carries a radiometric 640 × 512 thermal camera next to a visible camera and records each motor, bearing and flange from the same viewpoints on every round.
On one patrol, the system flags a pump bearing at 78 °C, about 25 °C hotter than the identical pump beside it. The maintenance team inspects the bearing the next morning, finds degraded lubricant and replaces it during a planned stop. The thermal patrol thus turns an unplanned shutdown into a routine repair and gives the plant a temperature history for every monitored asset.