Why edge inference
Detection has to run on the aircraft. The mission code needs answers immediately — track this, re-pass that, notify now — and a video round-trip to the ground station adds latency and a single point of failure. So the classifier lives next to the camera.
V1: thresholds, and an honest failure
The first pass was classic computer vision: threshold the thermal image, extract contours, call anything hot an animal. It worked beautifully indoors. Outdoors in a Texas summer it fell apart — by afternoon, the ground is hotter than the animals, and anything cooler than the background is invisible to a hot-spot detector. Good lesson, cheaply learned.
V2: YOLO on an accelerator
The replacement is a YOLOv5-nano model quantized to TFLite running on a Coral Edge TPU, with export paths also built for TensorRT and ONNX. Detection results were significantly better than thresholding from the first tests.
The interesting decision was architectural: Edge TPU + Raspberry Pi 4 versus consolidating everything on the Jetson Orin Nano. I wrote the tradeoff up properly — accuracy at altitude, power budget, and the maintenance cost of two computers versus one — and it remains the fork in the road when this resumes.
Status
Paused since the July 2025 flight tests. Next steps are unchanged: dataset collection flights, training on real captures instead of transfer-learning approximations, and wiring detections into the mission logic.