Autonomous Boat Vision

Term definitionscoming soon

Definitions for this piece will land here. Inline terms in the body will link into this panel once they are filled in.

/** Fill in when the case-study rewrite lands. Used by ArticlePage Term definitions. */ export const glossary = [];

A model that works in a notebook and a model that works on a boat are different artefacts. This one had to run on an NVIDIA Jetson, in daylight, on water, inside a power budget, with nothing to phone home to.

The constraint

Everything interesting about this project is a consequence of one fact: there is no uplink. Inference happens on the vessel or it does not happen. That rules out the comfortable answer — stream frames somewhere with a real GPU — and turns model size, quantisation and frame budget into design decisions rather than afterthoughts.

What I built

Detection over the boat's camera feed, running on the Jetson, producing a continuous picture of what is around the hull for the navigation stack to use.

Decisions worth defending

• Latency over accuracy at the margin. A detection that arrives two seconds late is not a worse detection, it is a hazard.
• Handle glare as a first-class case. Water reflects the sky directly into the lens for a good part of the day; treating that as an edge case means building something that only works in the morning.

Result

PLACEHOLDER — sustained FPS on the Jetson, the classes it detects reliably, and the conditions where it stops being trustworthy.