projects / runsight
RunSight
AI glasses for visually impaired runners.
RunSight watches the track through a single camera and talks a visually impaired runner through it, in real time, so they can train on a public track without someone running beside them.
- 84.23%
- Lane segmentation IoU
- 88.24%
- Person detection IoU
- ~10 FPS
- On-device inference
- 6×
- Faster after optimization
- Offline
- No internet needed
- RGB
- One camera, no LiDAR
the problem
For a visually impaired runner, a public track usually means bringing a guide. Someone who runs alongside, holds your pace, and calls out the lane and whatever is in the way. It works. It also means you can only run when they are free.
The tools that exist don't really close that gap. A white cane misses changes in the track surface. GPS drifts by several meters. The few purpose-built devices are heavy, die in a couple of hours, and cost more than most people will spend. Running is fast and unforgiving, and it needs something made for it.
“If my guide is busy, my training gets postponed.”
Dimas, a runner from our field interviews
the system
intelligence
RunSight-AI
Three vision models run side by side on the device, reading the lane, the people ahead, and how close they are.
hardware
RunSight-Wearable
Camera glasses and a vest. The vest carries a Raspberry Pi 5 with a Hailo AI accelerator, GPS, batteries, and earphones.
app
RunSight-Mobile
A companion app, built to work with a screen reader, for pairing the device, tuning the voice, looking back at your runs, and sharing them.
the ai
One camera, three models running at once on the device. Every frame, their outputs fold into one spoken call, once per frame.
Lane segmentation
IoU 84.23%U-Net · ResNet-18
Traces the lane lines, reads whether the lane runs straight or bends, and works out which of five regions across the track you're in.
Person detection
IoU 88.24%YOLOv8n
Spots the people on the track ahead, the thing a runner most needs warning about.
Depth estimation
Monocular RGBSc-Depth V3
Works out how far off each person is from one camera, so only the ones close enough to matter set off a warning.
how it guides
Every frame comes down to one spoken cue, paced to how fast you’re going. And when the system isn’t sure, it says nothing rather than risk a wrong call.
Straight
Clear lane, nothing ahead
Turn left / right
The lane curves
Move left / right
Someone ahead in your lane, or you've drifted off it
Slow down
Someone ahead and no open lane to move into
inside the device
Camera glasses
A USB camera clipped to lightweight glasses, seeing the track from the runner's eyes. It's the only sensor the AI needs.
Compute vest
A Raspberry Pi 5 and a Hailo AI HAT, with GPS and a pair of 18650 cells on a UPS, all in a 3D-printed case that rides in a running vest.
Earphones
Voice cues over Bluetooth. For someone who can't glance at a screen, sound is the obvious channel.
building the ai
The models learn from a dataset we shot ourselves. Hundreds of first-person frames from real running tracks, hand-labeled in Label Studio, then stretched with augmentation to cover the light and motion of an actual run. Once trained, we quantize them and compile them for the Hailo chip, which is where the speed-up comes from.
watch the pitch
gallery
where it stands
Tested in
Variable lighting · crowded tracks · different speeds
Safety system
Threshold logic · frame smoothing · conservative rules
Recognition
Global Ambassador 2026 · 1st Place + People's Choice, Indonesia 2025
Status
Still in development
RunSight is a prototype, and we're careful to call it that. It has been out on real tracks, used by the people it's for, and it keeps changing. Where we want to land is a version any visually impaired runner can pick up and use alone, on any track, with nobody running beside them.