A 30 cm drone, through the two cameras of a search-and-confirm rig I'm modelling. Against a bright sky the noise is small next to the drone's contrast; the hard part is picking out a smudge five pixels wide among eight million, and knowing it isn't a bird. The project asks whether a physics model, calibrated once for each detector, can predict when a trained detector manages that on a camera it was never tuned on. None of it is a result yet: the numbers are illustrative, and the air here is perfectly still, which flatters the confirmation camera most.

Monochrome sensor, 3.45 µm pixels, 1 ms exposure, 550 nm. A dark drone at a tenth of the sky's brightness, hovering. 20 km visibility, no turbulence.

← Research