SafeMotion

An Android app that detects falls using on-device ML inference and phone sensors, then sends an SMS alert with location to a configured caretaker.

The TensorFlow Lite model bundled in app/src/main/assets/model.tflite is produced by the ML pipeline repo: Elderly-Fall-Detection. That repo also outputs deploy_config.json (input scaler stats + tuned threshold), which TFLiteRunner reads at runtime so the model and the app stay in sync across retrains.

Download

• Google Play Store: Coming soon — link will be added once the listing goes live.
• GitHub Release: SafeMotion v1.0 — download the APK directly from the latest release.

How It Works

The app runs a foreground service that continuously monitors accelerometer and gyroscope data. When a potential fall is detected, it goes through a 4-gate validation pipeline to eliminate false positives:

1. Impact Gate — Peak accelerometer magnitude must exceed 25 m/s² (~2.5g)
2. Free-Fall Gate — A dip below 6 m/s² must precede the impact (the brief weightlessness before hitting the ground)
3. ML Gate — A TensorFlow Lite model must score the sensor window above the threshold loaded from assets/deploy_config.json (currently 0.45)
4. Stillness Gate — Post-impact accelerometer data must show the person is stationary for ~2.5 seconds (rules out trips where the person recovers)

Only when all four gates pass does the app trigger an alert.

Alert Flow

1. Phone vibrates and plays an alarm sound
2. A 15-second countdown gives the user time to cancel (false alarm)
3. If not cancelled, an SMS is sent to the caretaker with:
   • User's name and phone number
   • Google Maps link to last known location

Features

• Two monitoring modes:
  • Constant — Sensors always at high rate (~50 Hz). Most reliable.
  • Adaptive — Starts at low rate, switches to high rate on significant motion. Saves battery.
• Sliding window inference — Overlapping sensor windows ensure falls at window boundaries aren't missed
• 3-minute cooldown between alerts to prevent repeated firing
• Setup wizard for configuring user and caretaker details
• Test Alert button to verify SMS delivery without simulating a fall

Architecture

ForegroundSensorService    Main service — sensor listener, pipeline orchestrator, alert dispatch
├── SensorBuffer           Sliding window of accelerometer + gyroscope samples
├── FallValidator          Physics-based gates (free-fall, impact, stillness)
├── TFLiteRunner           TensorFlow Lite model inference
├── WakeDetector           Adaptive mode: gates LOW↔HIGH rate switching
├── MetricsLogger          Tracks sensor events and inference counts
└── AppStatusNotifier      Persistent status notification

Requirements

• Android 8.0+ (API 26)
• Permissions: SMS, Location, Notifications, Body Sensors

Building

./gradlew assembleDebug

The TFLite model file (model.tflite) must be placed in app/src/main/assets/.

Testing on Device

Monitor the detection pipeline in real-time:

adb logcat -s INFERENCE:D ALERT:D ADAPTIVE:D MODE:D

Log tag	What it shows
INFERENCE	Gate decisions, confidence scores, stillness results
ALERT	SMS sent/failed, cooldown, cancel events
ADAPTIVE	Rate switching (HIGH/LOW)
MODE	Service start mode (ADAPTIVE or CONSTANT)

Tech Stack

• Kotlin
• TensorFlow Lite 2.12
• Android Foreground Service
• SensorManager (Accelerometer + Gyroscope)
• SmsManager for alert delivery
• LocationManager for GPS coordinates