HAR Demo — CIS4930

Smartphone-based Human Activity Recognition. Train a 1D CNN on the UCI HAR dataset, ship a single-page PWA that runs inference in the browser, and share predictions live between devices via Supabase Realtime Broadcast.

See docs/ for the design docs.

Repo layout

training/   Python pipeline (download UCI HAR, train CNN + RF, export to TF.js)
web/        Static PWA (HTML/CSS/ES modules, TF.js from CDN, Supabase from CDN)
docs/       Design + brainstorming notes
vercel.json Static hosting config

Quick start

1. One-time Supabase setup (~2 minutes)

1. Create a Supabase project (free tier). https://supabase.com/dashboard
2. No schema, no tables — we use Realtime Broadcast only.
3. Copy the Project URL and anon public key from Project Settings → API.
4. Copy the config template and fill in your values:

   cp web/js/config.example.js web/js/config.js

   Edit web/js/config.js and paste your URL + anon key.

2. (Optional) Train the model

The PWA ships with a heuristic fallback (STATIONARY/MOVING) so it works end-to-end before the CNN is trained. To get the real 6-class model (requires Python 3.11 — TF 2.15 + tfjs 4.10 don't ship 3.12 wheels):

cd training
/opt/homebrew/bin/python3.11 -m venv .venv && source .venv/bin/activate  # brew install python@3.11 if needed
pip install -r requirements.txt
python train.py

This downloads UCI HAR, trains the CNN + RF baseline, and writes the TF.js model into web/model/. Next time you load the PWA it auto-detects the model and uses the CNN instead of the heuristic.

3. Run locally

Any static file server works. Sensor APIs require HTTPS except on localhost:

cd web
python3 -m http.server 8000
# open http://localhost:8000 on your laptop for testing

For phone testing you need HTTPS. Easiest: deploy to Vercel (below) and open the preview URL on your phone.

4. Deploy to Vercel

npm i -g vercel
cd web
vercel           # first time: link/create project
vercel --prod    # deploy

Vercel auto-detects the static site. vercel.json at the repo root sets the Permissions-Policy header for motion sensors.

How the demo works

• One URL. All devices opening it join the same Supabase Broadcast channel.
• Tracker role. Any device that taps Start tracking grants motion permission, runs the CNN on-device (no data leaves the phone), and broadcasts {label, confidence, ts} at ~2 Hz.
• Viewer role. Any other device on the URL auto-shows the latest prediction in a huge readable label. That's the projector/laptop view while the phone is in a pocket.
• Fallbacks. If Supabase is unreachable, local tracking still works. If motion permission fails, the UI explains what to do. If the CNN isn't deployed yet, the heuristic classifier runs.

Presentation day checklist

• Supabase config filled in, deployed to Vercel.
• Visit URL on presenter laptop (plugged into projector) — should show "WAITING".
• Teammate visits URL on phone, taps Start tracking, grants motion permission, pockets phone.
• Laptop should switch from "WAITING" to a live label.
• Backup: 60 s screen recording of a good run, embedded in slide deck.

Permissions + quirks

• iOS Safari requires a user gesture to request DeviceMotionEvent.requestPermission() — we call it from the Start tracking button.
• Wake Lock is requested while tracking so the screen doesn't sleep in a pocket.
• HTTPS required for sensor access. localhost is exempt for dev.

License / scope

Course project for CIS4930, Spring 2026. Not production software. Supabase anon key is public-safe (Broadcast only, no tables, no RLS) for the duration of the course.