← Maya ChenFall Detect
Spots falls from wrist accelerometer data, for seniors living alone in HDB flats. Year 2 team project with a small labelled dataset.
- Maya ChenFeature extraction and evaluation
- Daniel Lim ✓Data collection and labelling
- Nur Aisyah ✓Data collection, test protocol
- Phone app and live demoNot confirmed yet
- Commits
- 87
- Stars
- 9
- Last commit
- 8 months ago
My part
I built the feature extraction (sliding windows, the stillness-after-impact feature) and the evaluation script, and ran the weekly check-ins where we compared results. I also wrote the README section on what our accuracy does and doesn't prove.
The problem
Most fall alarms need the person to press a button, which is exactly what someone who has fallen often can't do.
Who it's for
Seniors living alone in HDB flats, and the children who check on them from across the island. One teammate's grandmother in Ang Mo Kio was the starting point.
Why I built it
It was a team project with a real person behind it, and a good excuse to learn signal processing properly.
What I built
A classifier that flags a likely fall from wrist accelerometer data in real time.
How it works
Sliding 4-second windows, hand-built features (peak acceleration, jerk, post-impact stillness), and a random-forest classifier. I wrote the feature extraction and the evaluation script; two teammates collected and labelled the data.
Hard parts and decisions
Telling a fall from sitting down hard on a plastic stool. Adding 'stillness after impact' as a feature cut those false alarms the most.
What happened
Works well on our own small dataset (12 people). That's too small to claim real-world accuracy, and the README says so.
What I'd do next
A larger, more varied dataset, and testing on an actual watch instead of a phone strapped to a wrist.