Model, privacy & limitations

What runs on your device

This is a browser adaptation of HumanFallDetection by Mohammad Taufeeque and Samad Koita, based on commit 0f9fa006. Its original trained two-layer, 48-unit LSTM weights are retained. Its five pose/motion features and fall-warning filter are ported to JavaScript.

The original Python OpenPifPaf pose estimator is replaced with Google MediaPipe Pose Landmarker Lite for browser compatibility. Pose confidence, bounding boxes, tracking, camera frame rates, and missing-body handling differ. Therefore this adaptation does not inherit any published accuracy result. It has not been validated for real-world safety monitoring.

What the overlays mean

Green means a body is tracked, not that the person is safe. Gray means warming up or insufficient pose information. Amber means the classifier is evaluating a possible fall. Red means a possible fall prediction passed the upstream-style temporal filter. Tracking IDs are temporary, not personal identities. Hidden limbs are not reconstructed.

The model has seven internal activity classes, with class zero treated as fall by the upstream application. Raw model outputs are not calibrated probabilities. The browser reports each tracked person separately, unlike the original first-person-only display logic. Tracks reset after interrupted or missing observations.

Camera & privacy

Camera access begins only after Enable camera and your browser’s permission prompt. No microphone access is requested. Images and model predictions are processed in browser memory, not uploaded, recorded, or stored. Model/runtime assets are served from this site; the hosting provider receives normal page/asset requests and may retain access logs, but this app sends no camera imagery or predictions.

Stop camera, leaving the page, or switching tabs releases the camera and ends detection. Your operating system may also suspend processing when the phone is locked or the browser is backgrounded. No background notifications or contact escalation are provided. A local alert tone is optional.

Use safely

Mount the camera still, disable automatic framing/zoom, and keep the full body and floor visible. Lighting, blankets, furniture, slow collapses, partial bodies, multiple people, and camera movement can cause missed detections or false alarms. A handheld phone aimed at someone is not a stable safety-monitoring setup. Never rely on this prototype instead of normal care or emergency systems. Safely staged tests should involve capable, consenting adults only.

Licenses & attribution

The upstream code and classifier are MIT licensed. Google MediaPipe runtime/model are distributed under Apache 2.0. See upstream MIT license, Apache license, and classifier provenance and tensor manifest.