Inspiration
A lot of us live far from our parents. Many Filipinos work abroad while their parents grow older at home, sometimes living alone. A fall is one of the scariest things that can happen to an elderly person, especially if nobody finds them for hours. Medical alert pendants help, but only if the person remembers to wear it and can still press the button. I wanted something that watches out for them without them needing to do anything.
What it does
AetherVision uses a camera to watch for falls and shows the status on a simple web dashboard for the family.
It tracks a person's posture using AI pose detection
If someone ends up lying down and stays still for a set time, it counts it as a fall
An alert with a single snapshot is sent to the dashboard
Family can check the live status and the alert history anytime
Privacy was a big deal for me. The raw video never leaves the camera device. Only the fall alert and one snapshot are sent out, so nobody is streaming video of their parent's home to the internet.
How we built it
It's split into two parts. The detector runs in Python on a small device near the camera, like a Raspberry Pi 5, an old laptop or a mini PC. It uses YOLO pose detection with Ultralytics and OpenCV, tracks each person, and runs a simple state machine that decides when something is really a fall. It works with a USB webcam, an IP camera stream, a video file or even a screen capture for demos.
The second part is a small Flask server that stores the events and serves the family dashboard, which refreshes every two seconds.
The heavy AI work can't run in a browser, which is why the vision part runs on the edge device and only the lightweight dashboard needs to be online.
Challenges we ran into
The hardest part is telling a real fall from normal things like sitting on the floor, lying on the couch or bending down. I added settings for how flat the body has to be, how long the person stays down and how still they are, so it can be tuned for each room. Furniture blocking the camera is the biggest problem, so mounting the camera high and angled down made a big difference.
Accomplishments that we're proud of
It works end to end with real cameras, and it respects privacy by design. It's also open source under the MIT license, so anyone can use it for their own family and improve it.
What we learned
AI detection is the easy part. Deciding when to raise an alarm without too many false alerts is the real work. I also learned a lot about running computer vision on small, cheap devices.
What's next for AetherVision
Instant Telegram alerts with a photo so family gets notified right away, a cloud dashboard with Supabase and Vercel so it can be checked from any phone, and support for multiple cameras around the house.
Log in or sign up for Devpost to join the conversation.