TypologySienna Chen
InputLightweight

Webcam pose tracking

The camera reads the whole body.

The camera watches the whole body and reports where the arms, hips and head are. Raise your arms and the scene reacts. It asks for camera permission, downloads a model, and works best a step back from the laptop.

Breaks when
camera permission prompt5-10MB model downloadjitter without smoothing
Also called
pose estimation, MoveNet, PoseLandmarker, skeleton tracking

Prompts

Body-driven particle scene
MediaPipe Tasks Vision PoseLandmarker (lite model, GPU delegate) in a Web Worker or on the main thread with runningMode VIDEO at 30 fps, mirrored. Smooth landmarks with a One Euro filter (minCutoff 1.0, beta 0.02). The wrists emit particles into a Three.js scene that shares the camera aspect; the hips set a global sway. Show a start button that requests the camera, a skeleton overlay toggle, and a fallback demo mode that replays recorded landmarks when no camera is granted.
Gesture from pose
Derive three events from pose landmarks: arms raised (both wrists above the nose for 300 ms), lean left/right (shoulder midpoint offset from hip midpoint), and a clap (wrist distance below 8% of shoulder width with a velocity spike). Debounce each and emit custom DOM events. Provide the thresholds as constants with comments.
Privacy-first setup
Wire pose tracking so that no video frame ever leaves the tab: process on-device, never draw the camera image (only landmarks), release the stream on page hide, and show a persistent “camera on” indicator. Document what to say in the permission prompt copy.

Related

Resources · 2