CountIn

CountIn counts people in real time from a webcam with TensorFlow.js. Everything runs in the browser: no server, no install, and the camera feed never leaves the machine. The project is paused for now, kept online as a reference.

How it works

• Detection. A COCO-SSD model loaded through TensorFlow.js finds people in each frame. Detection runs on every other frame to keep the UI responsive.
• Tracking. A hand-written centroid tracker matches new detections to existing tracks by distance, with a confidence gate against flickering false positives and a short position history to survive occlusion.
• Line crossing. You draw counting lines; for each track the system tests whether the segment between its previous and current position intersects a line, using a cross product with a direction sign for in versus out.
• Dashboard. Counts update live per line, with running occupancy and a flow chart.
• Hub and camera stations. Several phones or webcams pair with a central dashboard by QR code. Each station runs detection locally and sends only numbers over WebSocket to a FastAPI backend, never video.

What is interesting

• The tracker is from scratch, so every stage is debuggable: when counts drift, you can see which stage is at fault.
• Counting is geometry, not ML: a cross product with a sign.
• The privacy story is short: the camera permission is the only permission, the model is fetched once and cached, and it works offline after the page loads.

TensorFlow.js performance varies a lot across hardware, so processing every other frame and rendering to canvas instead of the DOM is what keeps it feeling real-time.

Status

Paused. The live demo is offline right now; run it locally from the repo with Docker Compose (Vite frontend, FastAPI and PostgreSQL backend). The detection pipeline is covered by end-to-end tests with Playwright and a synthetic camera feed.