Thermal camera
Thermal camera – definition
A thermal camera is an imaging sensor that measures infrared radiation emitted or reflected by objects and converts it into a two-dimensional image. In mobile robotics, the term usually refers to a long-wave infrared (LWIR) camera operating in the 8-14 µm spectral range, where ordinary objects near ambient temperature emit detectable thermal radiation. Thermal cameras are also called infrared cameras, but this term can include near-infrared devices that work differently and often require active illumination.
For an unmanned ground vehicle (UGV), a thermal camera provides information that is not available from RGB cameras, LiDAR, or depth cameras. It can detect temperature contrast in low visible light and, depending on conditions, through some haze, smoke, or visually uniform scenes. The output may be radiometric, where each pixel represents a calibrated temperature-related measurement, or non-radiometric, where pixel values represent relative thermal intensity only.
On Leo Rover and Raph Rover, a thermal camera is normally integrated as an external payload connected to the onboard compute unit through USB, Ethernet, MIPI CSI, or a vendor-specific interface, depending on the camera and onboard hardware. It can be used for thermal inspection, hotspot detection, human-presence research, agricultural monitoring, and multimodal perception. It does not provide obstacle range by itself, so autonomous navigation still requires sensors such as LiDAR, stereo/depth cameras, wheel odometry, IMU data, or GNSS/RTK where applicable.
Thermal imaging principle
Thermal imaging depends on infrared radiance rather than visible reflected light. A warmer object generally emits more thermal radiation than a cooler object, but apparent temperature is also affected by emissivity, reflected radiation, atmospheric transmission, distance, and the camera calibration model.
The radiant exitance of an ideal black body is described by the Stefan-Boltzmann law:
M = σT⁴In this expression, M is radiant exitance in W/m², σ is the Stefan-Boltzmann constant, and T is absolute temperature in kelvin. Real materials are not ideal black bodies. Their thermal emission is commonly modelled using emissivity ε, where 0 ≤ ε ≤ 1. Temperature measurements therefore require material-dependent emissivity settings and a calibrated radiometric camera.
Most UGV thermal payloads use one of the following detector classes:
- Uncooled microbolometer – common for LWIR inspection and mobile robotics. It measures changes in detector resistance caused by infrared heating.
- Cooled photon detector – used where high sensitivity, high frame rate, or mid-wave infrared imaging is required. It adds cost, power demand, cool-down time, and mechanical complexity.
- Thermopile array – low-resolution thermal sensing, typically used for simple temperature monitoring rather than navigation-grade imaging.
FLIR identifies 8-14 µm as the spectral range of its Boson+ LWIR camera modules. The manufacturer specifies available resolutions including 320 × 256 and 640 × 512 pixels, depending on the module variant. Source: FLIR Boson+ product specification.
Thermal camera data in ROS 2
In ROS 2, a thermal image is typically published using the same standard interfaces as a visible-light camera. The primary message type is sensor_msgs/msg/Image. Camera calibration data is published separately in sensor_msgs/msg/CameraInfo. These interfaces allow thermal images to be processed by ROS 2 image pipelines, synchronized with other sensors, and recorded in rosbag files.
A driver should publish timestamps from the acquisition source where possible. The image frame should follow ROS coordinate frame conventions defined by REP 103. For an optical camera frame, REP 103 specifies the conventional axes as x right, y down, and z forward. Source: ROS REP 103 – Standard Units of Measure and Coordinate Conventions.
| ROS 2 element | Typical value | Purpose |
|---|---|---|
| Image topic | /thermal/image_raw |
Raw thermal image stream |
| Message type | sensor_msgs/msg/Image |
Pixel matrix and timestamp |
| Calibration topic | /thermal/camera_info |
Intrinsic matrix and distortion model |
| Frame ID | thermal_camera_optical_frame |
Transform reference for fusion |
The ROS 2 Image definition includes image height, width, encoding, endianness, row step, and pixel data. Supported encoding names are maintained by the sensor_msgs package. A radiometric sensor may expose 16-bit single-channel values, often transported as mono16, while a false-colour visualization stream may use rgb8 or bgr8. The chosen encoding must be documented by the camera driver because a 16-bit value is not automatically equivalent to a temperature in degrees Celsius. Source: ROS 2 sensor_msgs/Image documentation.
ros2 topic list | grep thermal
ros2 topic echo /thermal/camera_info --once
ros2 run tf2_ros tf2_echo base_link thermal_camera_optical_frameKey parameters for UGV integration
Thermal camera selection should be based on measurable system parameters, not only the displayed image quality. A low-resolution sensor may be sufficient for hotspot detection at short range, while target classification or thermal mapping requires suitable optics, calibration, synchronization, and compute capacity.
- Detector resolution – number of thermal pixels, such as 320 × 256 or 640 × 512.
- Frame rate – acquisition frequency in frames per second. It affects motion blur, synchronization, and CPU or GPU load.
- Field of view – determined primarily by the lens. A wide lens improves nearby scene coverage but reduces pixels on distant targets.
- NETD – Noise Equivalent Temperature Difference, commonly expressed in mK. Lower NETD indicates the ability to distinguish smaller thermal contrasts under specified conditions.
- Radiometric accuracy – relevant only for calibrated temperature measurement. It must be evaluated against the manufacturer’s stated conditions, emissivity configuration, and environmental limits.
- Interface and bandwidth – USB, GigE, CSI, or analogue video affect driver availability, cable length, latency, and onboard compute integration.
Use with Leo Rover and Raph Rover
Leo Rover runs ROS 2 natively on its built-in Raspberry Pi compute unit and uses four-wheel differential drive. A thermal payload can be mounted with its optical axis aligned with the robot heading or angled toward the ground or inspected asset. The static transform between base_link and the thermal camera frame should be published through URDF or a static transform publisher. Leo Rover is IP55-rated, but the thermal camera, connectors, enclosure, and cable entries must be assessed separately for outdoor operation.
Raph Rover is suitable when the thermal sensor must be combined with a larger battery, protective enclosure, pan-tilt unit, LiDAR, GNSS/RTK receiver, or additional compute hardware. The thermal stream should be synchronized with navigation and perception data before it is used for georeferenced inspection or sensor fusion.
Thermal images are useful for detecting thermal anomalies, but they are weak as a standalone navigation modality. Flat surfaces with similar temperatures may contain little usable visual texture. Automatic gain control can also change image contrast between frames, which can reduce the consistency of image-based tracking. For this reason, thermal perception is commonly fused with LiDAR, RGB/depth imagery, wheel odometry, and IMU measurements rather than replacing them.