SafeGuard-AI: TinyML Fall Detection System
Real-time on-device fall detection deployed on Arduino UNO Q with quantized inference
System Architecture & Challenge
Falls are a major hazard for elderly individuals and industrial workers. Vision-based monitoring raises serious privacy concerns and suffers from occlusions, while cloud-connected sensors introduce network latency and depend on continuous connectivity.
SafeGuard-AI was engineered to operate completely disconnected from external networks. By shifting deep learning inference directly onto a low-cost, resource-constrained microcontroller (Arduino UNO Q), the system achieves deterministic, millisecond-scale on-device detection without streaming raw sensor telemetry off-device.
Technical Implementation
Key Contributions & Verified Scope
- Engineered motion features from raw 6-axis IMU (accelerometer & gyroscope) sensor data to distinguish dynamic fall signatures from normal activities of daily living.
- Trained a deep learning fall detection architecture optimized specifically for microcontroller deployment constraints.
- Quantized model weights using TensorFlow Lite post-training quantization, achieving real-time on-device execution on the Arduino UNO Q microcontroller platform.
- Awarded National Winner in the Qualcomm × Arduino 'Physical AI: AI for All' Innovation Challenge out of 100 participating teams across India.
- Selected as an Industry Exhibitor to demonstrate SafeGuard-AI live at Qualcomm Booth 4.2 during the India AI Impact Summit 2026 at Bharat Mandapam, New Delhi.
Demonstration & Summit Highlights
Photographs, award presentations, and media coverage from the India AI Impact Summit 2026 at Bharat Mandapam, New Delhi.
