Computer vision · Project notes
How the fall detector evaluates motion
The desktop prototype uses YOLO to estimate body pose, then compares changes across frames before it logs a fall-like event.
Start with body keypoints
The pose model estimates joints and confidence values from the selected camera frames. The prototype drops incomplete or unstable detections before comparing hip height, body angle, vertical movement, and recent posture.
Check whether the movement persists
The detector keeps a short history and uses configurable timing thresholds. It looks for several pieces of evidence across frames, then uses recovery and cooldown states to limit repeat alerts for the same event.
Render and log the event
A skeleton view can display the measured pose without showing the source video. The application can log events locally and optionally pass them to a notification adapter. The public demo is a recording of the desktop research prototype.