il prototipo
Il display in azione
circuito simulato
web interface

Inspiration

Pedestrian crossings are one of the most dangerous places in a city: pedestrians, cyclists and cars share the same space, and many accidents happen simply because someone notices the others *too late. We asked ourselves: what if the crossing itself could see what's coming and warn people *before a conflict happens?

That's how BeSafe was born.

What it does

BeSafe turns an ordinary crossing into a smart one. Using a camera and an Arduino UNO Q, it:

detects pedestrians, cyclists and vehicles in real time
estimates speed and trajectories to predict a possible conflict before it happens
triggers light and sound alerts when the situation becomes dangerous
counts crossings by direction, producing useful data for the city

All the AI runs locally on the board: no cloud and no video leaving the device, which is better for privacy and latency.

How we built it

Arduino UNO Q: the Linux side runs the Python computer-vision pipeline, while the microcontroller drives LEDs, buzzer and display.
Arduino App Lab: used to develop, connect and deploy both parts of the app.
Danger rule: for each detected object we estimate its velocity $v$ and its distance $d$ from the crossing area, and compute the time needed to reach it:

$$ t = \frac{d}{v} $$

If a vehicle's $t$ falls below a threshold $\tau$ while a pedestrian is crossing, the system raises an alert.

Real-time web dashboard (HTML, CSS, JavaScript, Socket.IO): it shows the live video with detections, the current danger state, adjustable thresholds and crossing counters.

We tested BeSafe on a real crossing in Turin.

Challenges we ran into

Running AI on the edge: getting detection fast enough on a small board meant balancing accuracy and speed.
From positions to predictions: turning noisy frame-by-frame detections into stable speed estimates was harder than expected.
Two worlds on one board: making the Python side and the microcontroller talk reliably in real time.
The real world: changing light, occlusions and unpredictable people behave very differently from a test video.

What we learned

How to build an end-to-end edge-AI system, from sensors to alerts.
How to design a simple but effective risk rule and tune its thresholds on real data.
How to work fast as a team under hackathon pressure, splitting hardware, AI and interface work.