🚗 Bantay Drive — Real-Time Driver Monitoring System

Undergraduate Thesis Project — New Era University, AY. 2025-2026

Authors: Macalanda, Pia Katleya V. & Mancera, Juliana R.

---

About

Bantay Drive is a mobile-based real-time driver monitoring system powered by an on-device deep learning model. It uses the front-facing camera to classify driver behavior into 13 sub-classes across three parent states (NATURAL, DISTRACTED, DROWSY), escalating tiered alerts before dangerous situations occur — entirely offline, 14.12 MB, no cloud dependency.

"DMS-HybridNet: A Dual-Stream Architecture Combining MobileNetV3 and Residual 1D-CNN for Real-Time Multi-Class Driver Behavior Monitoring on Mobile Edge Devices"

---

Authors

Name	Role
Pia Katleya V. Macalanda	Machine Learning Engineer, Dataset Preparation, UI/UX Design, Researcher
Juliana R. Mancera	Mobile App Developer, Model Integration, Testing & Deployment, Researcher

Institution: New Era University, College of Informatics and Computing Studies
Program: Bachelor of Science in Computer Science
Year: 2025-2026

---

License

Copyright © 2026 Pia Katleya V. Macalanda & Juliana R. Mancera. All rights reserved.

This project was developed as an undergraduate thesis at New Era University. It is licensed under CC BY-NC-ND 4.0 — you may view and cite this work with attribution, but you may not use it commercially, distribute modified versions, or deploy any part of it (including the DMS-HybridNet model) without prior written permission from the authors.

---

Features

Real-Time Monitoring

• On-device TFLite inference (NNAPI → CPU fallback) — no server required
• 3-level escalating alert system:
  • Level 1 — Slide-in banner; alert sound plays once (auto-dismisses)
  • Level 2 — Persistent banner; alert sound plays 3 times
  • Level 3 — Full-screen blocking alarm overlay; alarm loops continuously until manually dismissed
• Configurable alert sensitivity:

Sensitivity	L1	L2	L3
Low	5 frames	10 frames	15 frames
Medium	3 frames	6 frames	9 frames
High	2 frames	4 frames	6 frames

• Head-pose visual indicator — real-time circle overlay tracking driver head rotation
• Video clip capture — automatically records and saves clips (up to 10 s) when alerts trigger; disk-space-aware (requires 50 MB free) with structured error codes; clips exportable to device Downloads folder
• Picture-in-Picture (PiP) — monitoring continues in a floating window when app is backgrounded
• Foreground service with persistent notification showing live driver state + Stop button
• Clear Glasses toggle, Auto-start recording option

Dashboard

• Circular Safety Score (0–100), color-coded green / amber / red
• Stat cards: Total Drive Time, Alerts (last 24 h), Safety Streak, Avg Alertness
• Safety Score History line chart (last 30 days, horizontally scrollable)
• Shimmer skeleton loading while data is being fetched

Analytics

• Time filter: 7 Days / 30 Days / All Time
• Drowsiness vs. Distraction daily line chart + Hourly Alert Distribution bar chart
• Shimmer skeleton loading while analytics data is being fetched

History

• Chronological session list grouped by date with search and filter chips
• Advanced filtering for both session logs and video logs
• Session detail: state breakdown, alert events (L1/L2/L3), system log
• In-app session video playback (non-mirrored)

Settings

• Alert volume, sensitivity, auto-start recording, data retention (7 Days / 30 Days / Forever)
• Clear all history with confirmation

---

Architecture

lib/
├── core/
│   ├── database/
│   │   ├── database_helper.dart       # SQLite — 6 tables, schema v4, migrations
│   │   └── db_change_notifier.dart    # Riverpod reactive DB counter
│   ├── inference/
│   │   └── tflite_service.dart        # Model loading, inference, 13-class mapping
│   ├── preference/
│   │   └── preference_helper.dart    # SharedPreferences wrapper
│   ├── services/
│   │   ├── notifications.dart        # Foreground service + notification management
│   │   ├── head_pose_service.dart    # ML Kit head-pose & euler angle calculation
│   │   ├── pip_service.dart          # Picture-in-Picture control
│   │   └── video_clip_service.dart   # Alert video recording & clip management
│   ├── providers.dart                # Riverpod state providers
│   └── session_state.dart           # Session data container
├── screens/
│   ├── monitor_screen.dart           # Camera + inference + alerts + PiP + head-pose
│   ├── dashboard_screen.dart         # Safety score + charts + skeleton loading
│   ├── analytics_screen.dart         # Trend charts + skeleton loading
│   ├── history_screen.dart           # Session list + video playback + filters
│   ├── settings_screen.dart          # App settings
│   ├── onboarding_screen.dart        # First-launch walkthrough
│   └── splash_screen.dart
├── widgets/
│   ├── exit.dart                     # Exit confirmation dialog
│   └── head_pose_indicator.dart      # Camera alignment visual overlay
├── utils/
│   └── responsive.dart              # Breakpoints + brand-specific scaling
└── main.dart                         # App shell + IndexedStack + landscape sidebar

Local Database (SQLite — 6 tables, schema v4)

• sessions — drive sessions with timestamps, safety score, trip label
• state_counts — neutral / drowsy / distracted frame counts per session
• alert_events — alert type, level (1/2/3), timestamp
• system_logs — INFO / SUCCESS / WARNING log entries per session
• alertness_snapshots — 5-second alertness readings per session
• video_clips — saved alert clip paths, alert types, duration

---

Getting Started

Prerequisites

• Flutter SDK 3.16+
• Android Studio / VS Code
• Android device or emulator (API 26+, Android 8.0 Oreo minimum)
• JDK 21

New to Flutter? The official Flutter documentation offers tutorials, samples, and a full API reference. A guided first-app walkthrough is available at docs.flutter.dev/get-started/codelab.

Model & Asset Setup

Place the following under assets/:

assets/
├── model/
│   └── dms_hybridnet_v3_float32.tflite
├── norm_params.json
├── L1_L2_sound.mp3
├── L3_critical_alert.wav
├── car.png
├── text_logo.png
└── bantay_drive_logo.png

Verify pubspec.yaml declares all of these under flutter: assets:.

Android Gradle Setup

android/app/build.gradle.kts — Java 21 target:

compileOptions {
    isCoreLibraryDesugaringEnabled = true
    sourceCompatibility = JavaVersion.VERSION_21
    targetCompatibility = JavaVersion.VERSION_21
}
kotlinOptions { jvmTarget = "21" }

dependencies {
    coreLibraryDesugaring("com.android.tools:desugar_jdk_libs:2.1.4")
    implementation("androidx.concurrent:concurrent-futures:1.2.0")
    implementation("androidx.concurrent:concurrent-futures-ktx:1.2.0")
    implementation("androidx.multidex:multidex:2.0.1")
}

android/gradle/wrapper/gradle-wrapper.properties:

distributionUrl=https\://services.gradle.org/distributions/gradle-8.13-all.zip

android/app/build.gradle.kts — suppress .tflite compression:

aaptOptions {
    noCompress += listOf("tflite")
}

Release builds use ProGuard minification + resource shrinking by default.

---

Key Dependencies

Package	Purpose
flutter_riverpod	State management
sqflite	Local SQLite database
camera	Camera feed + image stream
tflite_flutter	On-device TFLite inference
google_mlkit_face_detection	Face detection for head-pose
flutter_foreground_task	Foreground service + persistent notification
fl_chart	Line + bar charts
shimmer	Skeleton loading animations
audioplayers	Alert sounds
volume_controller	System volume control
video_player	Session video playback
sensors_plus	Accelerometer (phone tilt)
shared_preferences	Settings persistence
permission_handler	Runtime permissions
device_info_plus	Brand-specific UI scaling
package_info_plus	App version display
path_provider	App documents directory
path	File path manipulation
url_launcher	Authors' GitHub links

---

On-Device Performance

• TFLite float32, 14.12 MB
• 5–15 ms per forward pass on mid-range Snapdragon
• Effective prediction rate: 6–7 FPS (every 5th camera frame)
• Privacy by design: no video transmitted, no cloud sync — compliant with the Philippine Data Privacy Act of 2012 (RA 10173)