After a several-month hiatus (moving houses interrupts flight testing more than weather does), the wildlife platform got back in the air.
Ground station progress. The Pelican case bottom panel is manufactured and mounted, and a repurposed Beelink mini PC now runs Ubuntu with the WFB-NG video downlink configured — the beginnings of the proper ground station.
The flight. Goals were video downlink stability and checking camera mount vibration. The FLIR is on a temporary 3D-printed mount for now; a gimbal comes later.
The detector reality check. The first-pass thermal pipeline used contour and threshold-based detection — find hot blobs, done. Indoors it worked great. On a hot summer day it fell apart: when the ground is hotter than the animal, “hot blob” stops meaning anything. Targets cooler than background are simply invisible to it.
The path forward. Moving to a learned detector. Currently testing YOLOv5 nano compiled to TFLite on an Edge TPU, with significantly better results already. The open question is compute architecture — the TPU rides the RPi4, but the Pi’s Ethernet is already dedicated to flight controller telemetry, which complicates cross-computer comms; mounting the Jetson instead would consolidate everything at the cost of a power and weight budget hit.
Next: video range testing, effective detection altitude, then the real dataset collection flights for classifier training.