Edge Fall-Detection Agent
In progressFall detection running on a Raspberry Pi, exposed as MCP tools so an agent can answer questions and adjust settings in natural language.
- Edge AI
- Raspberry Pi
- YOLOv8
- MQTT
- MCP
The problem
Fall detection research often stops at a model. This project turns my Master's thesis work into a working edge system that processes video locally and can be queried and configured through an agent.
Approach
- Extract body keypoints with YOLOv8 pose estimation on the Raspberry Pi
- Classify falls from the keypoints, based on my thesis approach
- Publish events over MQTT
- Expose tools such as get_current_status, get_recent_events, and set_sensitivity through an MCP server
- Store only keypoints and events, not raw video
Challenges
Real-time performance on the edge
Measuring FPS and latency, and comparing CPU-only inference with hardware acceleration.
Reducing false alarms
Requiring the fall posture to persist before raising an alert, and tuning the trade-off with missed detections.
Privacy by design
Processing locally and keeping no raw footage.
Honest evaluation
Reporting false positives and false negatives instead of only headline accuracy.
A research and demonstration project, not a medical or emergency-response product.