See the hazard
before it reaches you.

SafeVision reads your camera feed, spots people, vehicles and obstacles, and flags anything getting too close. All in your browser. Nothing leaves your device.

On-device Video is never uploaded.
Live Alerts update every frame.
Tunable Choose what counts as a hazard.

Camera is off.

Show boxes Ready

Settings

Saved on this device. Changes apply immediately.

Detection Minimum confidence Alert distance An object triggers an alert when it fills this much of the frame. Lower alerts sooner. Camera

Alerts Show detection boxes Spoken warnings Warning beep Vibrate (phones)

Objects that count as hazards

How SafeVision works

Look. Your camera frames are analysed by COCO-SSD, a pre-trained model running through TensorFlow.js.
Judge. The larger an object appears in the frame, the closer it probably is. Big, confident detections of hazard objects trigger an alert.
Warn. A red overlay marks the hazard, a banner names it, and optional voice, beep or vibration gets your attention.

Privacy

Video stays in your browser. Only the model files are downloaded, once, when you first start.

Limits

Distance is estimated from object size, not measured. Speed, direction and lighting are not considered. Expect false alarms and misses.

Not a safety device

This is a prototype. Never rely on it for driving, cycling or walking. Stay aware of your surroundings.