Platform · Vision
Seeing what AIS misses.
AIS (Automatic Identification System) is the collision avoidance standard on commercial vessels and most large yachts. It is good at what it does. It is completely blind to anything that does not carry an AIS transponder: kayaks, swimmers, fishing nets, debris, whales.
32°N Vision adds cameras and on-board AI to the safety layer. Inference runs on the hub — no cloud round-trip, no satellite connection required. When something is detected, the alarm fires in under a second.
The problem
Most of what can sink you is invisible to AIS.
AIS works by radio — every vessel transmits its position, course, and speed on a shared frequency. You plot them on the chartplotter and watch for collisions. That covers the vessels that transmit. Not the ones that do not.
Most collision risk at sea is not from ships. On coastal and offshore passages, the things I actually worry about are the ones I cannot see on the chartplotter: a lobster pot line across the bow at night, a container lost from a ship, a whale sleeping on the surface. AIS is silent on all of them. A camera-based detection system is not.
Detection categories
What the vision system looks for.
The detection model is trained on maritime data — not general object detection repurposed for boats. The target categories matter: the system should alarm for a swimmer in the fairway, not a wave crest.
Small craft without AIS
Kayaks, rowing boats, paddleboards, RIBs, tenders, fishing dinghies. The most common near-collision scenario in crowded anchorages and harbour approaches. AIS-invisible, vision-visible even in low light with the right camera.
People in the water
Swimmers, divers on the surface, person overboard. The detection event fires on the cockpit display and sounds the alarm. On a boat with an autopilot connected to the platform, the turn-back manoeuvre can be triggered automatically.
Floating debris
Shipping containers lost overboard are rare but catastrophic. Fishing gear — nets, lines, pot markers — is far more common and easier to detect. The system looks for anomalous floating objects ahead of the boat's track.
Large marine mammals
Whales at the surface are a genuine hazard on offshore passages, particularly in the Atlantic and Mediterranean. They do not appear on AIS, they do not manoeuvre, and they are hard to see from deck level at night. A bow camera with detection can give a few seconds of warning — enough to alter course.
Vessels without AIS transponders
Many smaller commercial vessels, fishing boats, and working craft are not required to carry AIS or run it reliably. Some run AIS on standby — transmitting intermittently. Vision covers the gap, particularly in coastal waters where these vessels are common.
Anchorage watch
At anchor, the camera watches the surrounding water for approaching vessels — useful in crowded anchorages where a dragging neighbour poses more risk than wind or sea state. The detection integrates with the anchor watch app: one alarm system, both radio and visual threats.
How it works
Inference on the boat. No cloud round-trip.
Safety decisions cannot depend on a satellite internet connection. The vision model runs on the hub — the Mac mini in the nav station. Every frame is processed locally. The alarm fires in under a second of detection. The cloud is never in the loop for safety-critical paths.
Apple Neural Engine on M-series
The Mac mini hub runs inference on Apple's Neural Engine — a dedicated neural processor on M-series chips that handles machine learning workloads without involving the main CPU or requiring a GPU. It processes a 1080p frame for object detection in well under 100 milliseconds on the M2 chip.
That speed is what makes real-time safety detection practical on a boat computer drawing 10 watts. It would not be practical on a Raspberry Pi 5 without active cooling and significant throttling — which is part of why I chose the Mac mini as the reference hub hardware.
No cloud in the safety path
The detection model runs entirely on the hub. Camera frames are never sent to a cloud service. The inference result — a detection event with a bounding box, confidence score, and class label — goes onto the 32°N bus and triggers alarms through the alarm service. That all happens locally.
Detection events are logged and can be synced to BoatHub when there is connectivity — for review, for training data improvements, and for the record. But that sync is async and post-hoc. The alarm path never waits for it.
The pipeline
From frame to alarm in four steps.
The vision pipeline is simple: a camera generates frames, the hub runs inference, a detection event goes onto the bus, the alarm service fires. Each step is a normal platform component — not a special vision-only system.
Camera ingest
One or more cameras connected to the hub over USB (or network RTSP stream). The vision service reads frames at the configured rate — typically 4 to 10 fps for detection; full 30 fps for recording.
On-device inference
Each frame is passed to the detection model running on the Neural Engine. The model returns bounding boxes with confidence scores and class labels. Frames below the confidence threshold are discarded without generating an event.
Bus event
Detections above threshold emit a typed bus event — class, confidence, bounding box, camera ID, timestamp. The bus is the same one every other 32°N app uses. No vision-specific wiring.
Alarm and log
The alarm service subscribes to vision detection events. It fires on the cockpit display, pushes a notification to connected phones, and logs the event with a short video clip to BoatHub.
Status and roadmap
This is coming. Here is the order.
Vision is not in v1. It requires a trained maritime detection model and real-world testing before I am willing to ship it as a safety feature. That takes time to do properly. Here is the plan.
Camera ingest and recording
Camera input to the hub, frame storage to BoatHub, basic motion detection (not semantic object detection). Useful for anchorage watch and post-incident review. Ships as a tool in v1.5; the detection model follows.
Object detection model: small craft and debris
The first semantic detection categories: small craft without AIS, and floating debris. These are the highest-frequency scenarios and have the best available training data. Alarm integration with the anchor watch and collision alarm apps.
People in the water, marine mammals, full alarm integration
Person-in-water detection is the hardest category: high false- positive risk, highest consequence if missed. It requires more training data and more careful threshold tuning before I am comfortable shipping it as a safety feature. Marine mammal detection follows once the base model is solid.
Explore the rest of the platform.
Vision is one part of the safety layer. The full platform covers hardware, bus, OS, apps, and cloud.