The problem
iPBMS monitors patients through ordinary IP cameras. Care staff need to know within seconds when someone falls — but bending, lying down, sleeping, partial occlusion and camera-angle shifts all look like falls to a naive detector, and every false alarm erodes trust in the system.
Detection pipeline
- YOLOv8-Pose + MoveNet Thunder with joint-angle analysis over 17-keypoint skeleton tracking and temporal confidence filtering.
- Over 85% fall-detection accuracy with under 2 s from fall to alert under real monitoring conditions.
Making it fast
- Quantized YOLOv8-Pose to FP16 with TensorRT: inference 45 → 12 ms per frame on an RTX 3060.
- Moved post-processing and NMS to C++, off the Python hot path.
- Capacity grew from 2 to 6 concurrent streams per node.
Cutting false alarms
- ByteTrack IDs and a per-track state machine.
- A hip-keypoint velocity threshold to separate falls from lying down.
- 5-frame temporal validation gating Normal / Warning / Danger states, delivered within 3 s over WebSocket and Supabase Realtime.
Explainable incident reports
Keyframe selection, image captioning and machine translation produce multilingual incident reports with priority levels, delivered to care staff through Firebase Cloud Messaging.
In production
The multi-camera inference stack runs on-premise (RTX 3060 12 GB, Core i5, 32 GB RAM, Ubuntu 22.04), live on two production cameras, with weekly stakeholder reviews.