At night, the lookout's job changes shape. Hull silhouette disappears; what remains is a grammar of light — a white masthead light's height tells length, the arc of a red or green sidelight tells heading, a tow light or a trawl light tells class, a strobe or a pair of flashing blues tells trouble. COLREGS Rules 20–31 encode an entire vessel registry into lamp colours, positions and arcs.
An autonomous ship's camera must do the same reading. The published work on navigation-light detection and classification is thin — a subfield of a subfield — and honest about why it is hard: the signal is tiny, the background is lit, and the label is a function of what the light <em>means</em>, not just what colour it is.
Key findings
- Navigation-light recognition is a spatial-and-semantic problem: a light's meaning depends on its colour, vertical position, arc and company — not colour alone. Established
- Background city and harbour light pollution is the dominant confuser — the discriminating feature is temporal (the light is attached to a moving target) as much as spatial. Research demonstration
- Published work is largely prototype-scale on limited or private data; no benchmark like SeaShips exists for light classification. Established
- COLREGS compliance needs the light's <em>meaning</em> — power-driven vs fishing vs towed vs not-under-command — which is a classification task on a handful of pixels. Maritime-GPT analysis
- Night is not the camera's worst case (the light is the target), but range, occlusion and weather make reliable classification unsolved. Research demonstration
The grammar the machine must learn
A trained lookout reads lights compositionally: two white masthead lights in a vertical line, plus red and green sidelights, reads as a power-driven vessel underway, over 50m; sidelights arranged to indicate aspect reveal whether it's crossing, head-on or overtaking; a yellow tow light or the all-round whites of a fishing vessel change the rules that apply. The model must reproduce that grammar from a few saturated pixels on a moving, rolling image — a task that is part detection, part temporal reasoning, part rule-lookup. Established
Why the literature is thin
- <strong>Data scarcity.</strong> Public night-vision maritime datasets are rare; annotated light configurations rarer. Most published work is on small private corpora. Established
- <strong>Class imbalance is hostile.</strong> The classes that change the manoeuvre (trawler, not-under-command, constrained-by-draught) are the rarest in any dataset. Maritime-GPT analysis
- <strong>Signal-to-background.</strong> Harbour lights, shore sodium lamps, and fishing-boat floods create hundreds of false sources a ship-borne camera must reject every frame. Research demonstration
- <strong>Range.</strong> At COLREGS-relevant night range the light is sub-pixel to a few pixels — classification happens on a speck. Research demonstration
What works
The consistent finding across the small literature: <em>association beats recognition</em>. A light that moves with a radar/AIS contact, in the right bearing sector, is easy to attribute; an isolated speck against the city is not. The practical architectures therefore detect lights loosely, then <strong>associate</strong> them to the fused contact the radar already tracks — the camera confirms the target's aspect and class, it doesn't discover it. This mirrors the larger lesson of the whole vision stack: the camera is strongest as the fusion's confirmer, weakest as its primary detector. Maritime-GPT analysis
The safety-relevant subset
The subset that matters is small and well-defined: the lights that change the rules. Not-under-command, restricted-in-ability-to-manoeuvre, constrained-by-draught, fishing-gear-out — these are low-frequency, high-consequence classes. A system that classifies 95% of lights correctly but misses the one NUC light in a thousand frames has not solved the problem COLREGS is asking. The research has not yet published a reliable-rate estimate at that rarity. Maritime-GPT analysis
What we know
- Navigation-light reading is a compositional COLREGS task, not pure colour classification.
- Light-to-contact association against radar/AIS is the proven approach.
- The high-consequence classes are the rarest in training data — the known failure mode.
What we don't
- A published reliable-recognition rate at the rare, rules-changing classes.
- Whether temporal/multi-frame reasoning beats single-frame classification at range.
- How camera-only systems handle lights beyond the reach of the fused radar contact.
Being tested
- Video-sequence models that use light persistence and motion to disambiguate.
- Synthetic night-scene generation for rare-class augmentation.
- Association-first architectures: radar finds, camera reads.
Maritime-GPT analysis
Night is where the vision stack most earns its keep — and where it most needs the fusion. The light itself is bright; the problem is telling a trawler from a tug at three pixels against a city. The architecture that works is the one the research is converging on everywhere: radar and AIS find the target, the camera reads it. A camera that must discover and classify lights alone at night is doing the lookout's hardest job with the thinnest tools. Maritime-GPT analysis
This is a research feature, not operational guidance. Simulation results are reported as simulation; demonstrations are not deployments; company claims are attributed as such. Verify safety-critical decisions against official nautical publications and flag-state requirements.
Sources & evidence
The lighting code the classifier is being asked to read.
Documents the night/IR vision problem within the broader maritime EO field.
The small-target regime — navigation lights are the extreme case of small-object detection.
The main public source with paired day/night maritime video including NIR.