AIS is a miracle of cooperative safety — every compliant vessel announcing who it is, where it is, where it's going. Its weakness is the word 'compliant': the fishing boat, the leisure craft, the small vessel that either doesn't transmit or transmits wrong is invisible to the AIS-only picture. The camera is the opposite — it sees anything in frame but knows nothing about it.
Together they cover each other's blindness. The camera finds the unreported; the AIS disambiguates the reported. That pairing — vision for the sensor-blind, AIS for the sensor-honest — is the fusion the avoidance literature converges on.
Key findings
- AIS and camera are complementary by construction — one reports cooperative contacts, one detects uncooperative ones. Established
- The fusion problem is <em>association</em>: matching a camera blob to an AIS contact, and flagging camera-only contacts (the dark targets) and AIS-only contacts (the occluded ones). Established
- The camera's contribution in good conditions is classification and confirmation; the AIS's contribution is identity and intent-broadcast. Research demonstration
- The failure mode that matters is the AIS-dark small vessel in the camera's blind spot — the case neither sensor catches alone. Maritime-GPT analysis
- Camera+AIS fusion is the most operationally proven pairing — it's what ECDIS overlays already do visually, and the ML version formalises it. Established
The complementary blindness
| Sensor | Sees | Blind to | The cover |
|---|---|---|---|
| AIS | Identity, position, course of compliant vessels | Dark vessels, wrong/failed transponders | The camera finds the unreported |
| Camera | Anything in frame — size, class, aspect | Identity, intent, beyond-horizon | AIS names and projects the reported |
| Fused | Reported + unreported, classified + identified | Camera blind spot + AIS-dark overlap | The residual case the human covers |
What association actually does
The fusion's core task is matching: this radar return, this camera blob, this AIS contact — are they the same ship? When they agree, the contact is confirmed and classified. When the camera sees something AIS doesn't, it's a dark target — the alarm. When AIS reports something neither sensor sees, it's beyond line-of-sight or occluded — the warning. The three-way reconciliation is the perception system's primary output. Research demonstration
Where it fails
- <strong>The unreported + occluded.</strong> A small craft behind a wave, no AIS, no camera lock — the case that needs the human lookout. Maritime-GPT analysis
- <strong>Association confusion.</strong> Two contacts close together can be mis-assigned — the camera reads ship A, the AIS labels it ship B, the fusion merges wrong. Research demonstration
- <strong>Spoofed cooperative contacts.</strong> An AIS lie that the camera can't disconfirm — the adversarial case. Maritime-GPT analysis
- <strong>Small-vessel detection limits.</strong> The camera's known weakness (small craft at range) is the fusion's residual — AIS can't help where the camera can't see. Established
Why this is the operational pattern
ECDIS already displays AIS contacts overlaid on the chart; radar already shows raw contacts beside them. The ML fusion formalises what the officer does by eye — reconcile the layers. The avoidance systems that work don't add a new sensor; they formalise the reconciliation the human already performs, and automate the 'this doesn't match' flag. Maritime-GPT analysis
What we know
- CV+AIS fusion is the perception architecture the avoidance literature standardised on.
- Association — matching sensor contacts to AIS reports — is the core technical task.
- The residual risk is the unreported-and-unseen case.
What we don't
- Reliable association performance at range in dense traffic.
- The false-alarm budget for dark-target flags in congested water.
- Whether a fused system can be certified as the lookout for Rule 5.
Being tested
- Real-time association engines on bridge hardware.
- Dark-target detection and alerting pipelines.
- Fusion that explicitly surfaces disagreement rather than hiding it.
Maritime-GPT analysis
CV+AIS is the honest perception answer because each sensor's failure is the other's coverage — the camera finds the ship that doesn't announce itself, AIS names the ship the camera can't classify. The limitation to internalise is the overlap: a dark vessel in the camera's blind spot defeats both, which is why the lookout remains and why the fusion's correct output is the disagreement flag, not a confident picture. 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 sensor-fusion literature the pairing is built on.