[AfA2026-PhysicalAI] - SafeGuard-AI

Arduino® UNO™ Q 2GB

Modulino™ Movement

Power Bank

Power: USB-C Cable

VS Code

Personal Computer

Twilio API for WhatsApp

Arduino App Lab

TensorFlow

Qwiic Cables Bundle

<html><body>Elderly falls are a critical medical emergency where every second counts. Imagine a system that doesn't just detect an accident but immediately calls for help, even if the user is unconscious. SafeGuard AI transforms the Arduino UNO Q into a life-saving biometric monitor. Instead of relying on simple, error-prone threshold sensors, this project uses the power of the UNO Q to run a TinyML neural network locally on the device. It continuously analyzes movement patterns in real-time, distinguishing between daily activities and dangerous falls. When a fall is confirmed, the device springs into action: it locks the dashboard into an emergency state, sounds an alarm, and—if not cancelled—uses the Twilio API to make an automated voice call to a caregiver, reading out vital stats like impact force and skin temperature. <figure><img src="https://projects.arduinocontent.cc/43664a67-e671-49ba-94fd-d49349240b47.png"/><figcaption>This is the architecture of SafeGuard AI - Fall Detection.</figcaption></figure>Hardware Lineup The Arduino UNO Q acts as the intelligent central hub, powering both the sensor acquisition and the AI inference. Here is the full hardware lineup: <ol><li>Motion Input → Modulino Movement . This sensor contains the LSM6DSOX 6-axis accelerometer and gyroscope. It connects via Qwiic/I2C and feeds high-frequency motion data (100Hz) to the board. </li></ol><ol><li>Connectivity → The Arduino UNO Q's built-in WiFi is used to host the Web Dashboard and communicate with the Twilio Cloud API for emergency calls. </li></ol><ol><li>Power → A portable LiPo battery or USB-C power bank makes the device fully wearable. </li></ol> User Interface &amp; Feedback The application hosts a medical-grade Web Dashboard directly on the Arduino UNO Q, accessible from any phone or computer on the network. It features three main states: <ol><li>Live Monitoring: Shows real-time graphs of the accelerometer data and the user's current "Protection Profile" (e.g., Active vs. Sedentary). </li></ol><ol><li>Smart Calibration: A dedicated mode that "learns" the user's walking style over 15 seconds to automatically adjust the AI's sensitivity, preventing false alarms for active users. (15 second time-span was specifically chosen for demonstration, but for end product we have to calibrate it for some hours). </li></ol><ol><li>Emergency Overlay: When a fall is detected, the screen turns red with a "FALL DETECTED" warning. It displays a 10-second countdown, giving the user a chance to cancel the alarm before the emergency call is triggered. </li></ol><div id="gallery"><img id="gallery" src="https://projects.arduinocontent.cc/8e0de248-ad69-4edf-b9c7-c09983c65b5a.png"/><figcaption>Web UI Dashboard</figcaption>,<img id="gallery" src="https://projects.arduinocontent.cc/cb31f7a7-44a0-4242-8aaa-abdd14c55654.png"/><figcaption>Real-Time Sensor Data</figcaption>,<img id="gallery" src="https://projects.arduinocontent.cc/c5d76762-492e-44ee-b6bb-19cf3a814829.png"/><figcaption>System Activity Log</figcaption>,<img id="gallery" src="https://projects.arduinocontent.cc/3d9e136a-b497-45ad-beb9-0da5699ab206.png"/><figcaption>Patient Incident History</figcaption></div> The AI Model The core of the project is a custom TensorFlow Lite (TFLite) model trained specifically for fall detection. <ol><li>Training Pipeline: We built a custom training pipeline (included in the tinyml folder of the repo) to extract features like Jerk, Signal Magnitude, and Standard Deviation from raw sensor logs. </li></ol><ol><li>Edge Inference: The model is quantized to <code class="code">int8 </code>to run efficiently on the Arduino UNO Q's hardware. </li></ol><ol><li>Local Processing: By running the inference locally on the board, we ensure user privacy and zero latency—the device detects the fall instantly without needing to send raw sensor data to the cloud. </li></ol> Software Architecture The software is developed entirely within Arduino App Lab, leveraging the board's heterogeneous dual-core architecture to split high-speed sensing from high-level logic. The responsibilities are split as follows: <ol><li>MCU Core (C++ / Sketch): Runs the <code class="code">sketch.ino </code>firmware. It handles the "hard real-time" task of polling the LSM6DSOX sensor at a strict 100Hz and transmitting the data via RPC. </li></ol><ol><li>Processor Core (Python / Main): Runs the <code class="code">main.py </code>script. It manages the data buffer, runs the TFLite inference engine, hosts the Web UI, and handles the Twilio API integration. </li></ol><ol><li>Communication: The two cores exchange data seamlessly using the <code class="code">Arduino_RouterBridge </code>library. </li></ol> <figure><iframe src="https://youtu.be/_tpTryW2WmI"></iframe></figure></body></html>

SafeGuard AI

[AfA2026-PhysicalAI] - SafeGuard-AI

An intelligent fall detection system running on Arduino UNO Q that uses TinyML to detect accidents in real-time and triggers automated emergency calls.