A real-time fall detection system for elderly care and solo-living safety. Built as a final year project, Guardia uses pose estimation to watch over people who live alone — and acts fast when something goes wrong.
Guardia runs on a standard webcam and processes video in real time. When it detects a fall, it:
- Speaks an "Are you okay?" prompt through the device speakers
- Sounds an audio alarm if there's no response
- Sends an SMS and WhatsApp alert to a designated emergency contact
- Uploads a snapshot of the fall to cloud storage for reference
The whole pipeline — from fall to SMS delivery — takes around 12 seconds.
Rather than training a custom model from scratch, Guardia uses YOLOv8n-pose (a pre-trained skeleton detection model) and applies multi-signal heuristics on top of the output:
- Velocity spike — sudden downward acceleration of the body center
- Horizontal posture — aspect ratio of the bounding box flips when someone falls
- Spine angle — skeleton orientation crosses a threshold toward horizontal
A fall is only flagged when at least two of these three signals trigger simultaneously. This cuts false positives from simple actions like deep bending or sitting cross-legged.
Each person in frame is tracked independently. There's a 3.5-second grace period after a fall is detected — if the person gets back up on their own, no alert is sent. If they raise a hand and hold it up for a sustained period, they can cancel the alert themselves.
Tested on 75 video clips across 7 categories with 3 subjects:
Performance drops in backlit conditions (79%) and on slow sideways falls near walls. Full breakdown in EVALUATION.md.
Requirements: Python 3.9+, a webcam, and a Twilio account for SMS alerts.
git clone <repo-url>
cd guardia_final-year
python -m venv venv
source venv/bin/activate # Windows: venv\Scripts\activate
pip install -r requirements.txtOpen guardia/config.py and fill in your emergency contact number and Twilio credentials. Then run:
python main.pyThe YOLOv8n-pose model downloads automatically on first run (~6MB). You can also drop a pre-downloaded yolov8n-pose.pt into the models/ folder.
All thresholds live in guardia/config.py. The defaults work well in most indoor conditions, but you can tune them:
FALLING_VELOCITY_THRESHOLD = 0.12 # Higher = less sensitive to velocity spikes
RECOVERY_TIMEOUT_SECONDS = 3.5 # Grace period before escalation
INACTIVITY_THRESHOLD = 8 # Seconds motionless before alert triggers
ENABLE_SMS = True # Toggle Twilio alerts on/offguardia_final-year/
├── main.py # Entry point and webcam loop
├── requirements.txt
├── guardia/
│ ├── detector.py # Core heuristics and per-person state machine
│ ├── alerts.py # TTS, audio, SMS, WhatsApp, image upload
│ ├── config.py # Thresholds and credentials
│ └── night_mode.py # Low-light brightness enhancement
├── guardia-android/ # Android app (Kotlin + TFLite)
├── models/ # YOLOv8n-pose weights
├── logs/ # Alert history and fall snapshots
├── EVALUATION.md # Full test report
└── ANDROID_INTEGRATION_PLAN.md
There's a companion Android app under guardia-android/ that runs inference on-device using TensorFlow Lite and CameraX. It's functional but still being polished — setup steps are in ANDROID_INTEGRATION_PLAN.md.
- Works best with a camera mounted at roughly 1.5m height at a 45° downward angle
- Detection degrades in backlit scenes or heavy occlusion
- Slow sideways falls (where the person slides rather than drops) are sometimes missed
- Multi-person tracking is supported but hasn't been stress-tested beyond 2–3 people in frame
Python · OpenCV · YOLOv8 (Ultralytics) · pyttsx3 · Twilio · ImgBB · Kotlin · TensorFlow Lite · CameraX
Nipun Sujesh · Kripa Sekar . Anjitha J