Thermal imaging is essential for operational environments where RGB sensors fail, but selecting the right thermal core for edge AI requires a deep understanding of sensor physics and algorithmic constraints.
The thermal sensor market spans microbolometer cores from multiple manufacturers, with specifications that vary in ways that are not always intuitive. Making the wrong selection — optimized for image quality rather than edge AI inference performance — results in a sensor architecture that software optimization cannot correct.
Core Sensor Parameters
Noise Equivalent Temperature Difference (NETD) — NETD measures the smallest temperature difference the sensor can resolve against its noise floor. For edge AI perception models, NETD determines the effective contrast available between the target and its background. A sensor with poor NETD (high value, typically above 50mK) produces noisy imagery that degrades model accuracy, particularly for small or partially occluded targets at range.
Pixel Pitch — Pixel pitch (measured in microns) determines the physical size of the focal plane array and, combined with lens focal length, defines the instantaneous field of view and angular resolution. Smaller pitch enables smaller optics for the same resolution, which is advantageous under SWaP constraints. However, smaller pitch reduces the energy collected per pixel, requiring longer integration times that limit frame rate in dynamic scenes.
Detector Format — The detector resolution (320x240, 640x480, 1024x768) determines the field of view achievable with a given lens at a given pixel pitch. For detection at range, sufficient angular resolution to place enough pixels on the target is critical. A target that subtends fewer than 10 pixels in the image plane is typically below the reliable detection threshold for current generation models.
LWIR vs. MWIR for Edge AI
Long-wave infrared (LWIR, 8-14 micron) sensors use uncooled microbolometer technology. They are compact, low power, and require no cryogenic cooling — the practical choice for SWaP-constrained edge deployments. Their limitation is sensitivity: LWIR sensors cannot resolve the fine temperature differences that MWIR sensors can, limiting performance against low-contrast targets at range.
Mid-wave infrared (MWIR, 3-5 micron) sensors use cooled detectors that achieve far higher sensitivity and resolution. They are larger, heavier, power-hungry, and significantly more expensive. For high-performance defense applications where SWaP budget allows, MWIR delivers detection and classification performance that LWIR cannot match.
Sensor-Algorithm Co-Design
Edge AI imposes constraints on sensor selection that pure imaging applications do not face. Frame rate must match the inference pipeline throughput. The data interface must be compatible with the edge processor. Sensor control must support dynamic integration time adjustment without introducing frame drops during scene adaptation.
The correct approach is to select the sensor and the edge processor together, evaluating the complete sensor-to-inference pipeline as a system rather than selecting components independently and integrating them afterward. Sensor specifications that look good on a datasheet can create insurmountable inference pipeline problems when the hardware integration is attempted.