How Echo 1 detects obstacles at 30 FPS on a wearable

The core promise of Echo 1 is simple to state and hard to engineer: notice hazards in front of a walking user and say something useful about them, fast, on battery power, with no internet.

The pipeline

A ~200° wide-angle camera feeds frames to YOLOv8n running on the Hailo AI HAT+ (40 TOPS). Inference takes about 17 ms per frame, which sustains 30 FPS with headroom for the rest of the safety engine.

• Detection — YOLOv8n names objects from the COCO set: person, chair, car, bicycle, and so on.
• Proximity — estimated from the bounding-box area fraction; a person filling a third of the frame is close.
• Direction — the field of view is mapped to clock positions, so a hazard on the right front becomes '2 o'clock'.
• Arbitration — only one alert is emitted per direction, and the most severe hazard wins.

Why we say less, not more

Early prototypes narrated everything and users tuned them out within minutes. The alert manager applies cooldowns per hazard and per direction, so the device speaks only when something changed and matters. A spoken alert sounds like: 'Person, 12 o'clock, close.'

Why an accelerator instead of the CPU

On the Pi 5's CPU alone, YOLOv8n manages a few frames per second at high power draw. Offloading to the Hailo-10H brings 30 FPS at lower total power — the difference between a demo and a device someone wears all day.