Edge-AI Driver Vigilance and Risk Monitoring System with Fleet Supervision

Group 21 – Team Invictus

Team

E-Number Name Email
E/22/336 Dileka Sadaruwan e22336@eng.pdn.ac.lk
E/22/396 Pahan Prabash e22396@eng.pdn.ac.lk
E/22/188 Dhanushka Kavindaya e22188@eng.pdn.ac.lk
E/22/372 Imasha Sewmini e22372@eng.pdn.ac.lk

Table of Contents

  1. Introduction
  2. Solution Architecture
  3. Hardware and Software Designs
  4. Testing
  5. Detailed Budget
  6. Conclusion
  7. Links

Introduction

Driver drowsiness and distraction are important road-safety concerns, particularly during long-distance journeys and repetitive driving. Reduced vigilance may be associated with prolonged eye closure, head drooping, or sustained attention away from the road. Fleet operators may also lack an affordable way to review safety events across multiple vehicles.

This project proposes a low-cost, retrofit driver-vigilance and risk-monitoring system combining Edge AI, embedded systems, and a web-based fleet-supervision platform. A driver-facing camera provides visual input to a Raspberry Pi, which processes frames locally to identify potential drowsiness and distraction. An ESP32 interfaces with contextual sensors and controls local warning devices such as a buzzer, vibration motor, and warning LEDs.

Contextual information, including GPS-derived vehicle speed, may contribute to explainable driving-risk levels. A secure multi-user web application is planned for drivers, fleet managers, and administrators. Facial video is intended to be processed locally rather than continuously uploaded, and local detection and alerts are designed to continue during Internet outages.

The system is intended for transport operators, logistics companies, and other organizations operating vehicle fleets. It is a driver-assistance prototype and will not directly control the vehicle.

Objectives

Solution Architecture

The planned system combines visual monitoring, edge processing, embedded alerts, backend services, and a multi-user fleet-supervision web application.

High-Level Architecture

flowchart TD
    A[Driver-Facing Camera] --> B[Raspberry Pi]
    B --> C[Computer Vision and Temporal Analysis]
    C --> D[Driving-Risk Assessment]
    D --> E[Local Alert Command]
    E --> F[ESP32]
    G[GPS and Contextual Sensors] --> F
    F --> H[Buzzer, Vibration Motor and LEDs]
    D --> I[MQTT / HTTP]
    I --> J[Backend API]
    J --> K[(Database)]
    K --> L[Driver Dashboard]
    K --> M[Fleet Manager Dashboard]
    K --> N[Administrator Dashboard]

Main Components

The local detection-and-alert process is designed to continue during Internet outages. Events can be queued locally and synchronized when connectivity returns.

Hardware and Software Designs

Planned Hardware

Component Purpose
Raspberry Pi Edge computing and local computer vision
Driver-facing camera Capturing driver visual information
ESP32 development board Sensor interfacing and alert control
GPS module Estimating vehicle speed when available
Buzzer Audible warning
Vibration motor Physical alert
Warning LEDs Visual status and warning indication
Power supply, wiring, and enclosure Prototype integration

The final component list may change based on availability, feasibility, and budget.

Planned Software

User Roles

Role Main Functions
Driver View driving status, receive alerts, review trips, and check assigned vehicle/device status
Fleet Manager Supervise multiple drivers and vehicles, review trips and safety events, identify potential speed violations, and view summary reports
Administrator Manage user accounts, permissions, driver and vehicle records, devices, and system logs

Privacy and Security

The design aims to process facial video locally rather than continuously upload it to the server. Authentication, role-based access controls, device authentication, and secure server communication are planned.

Testing

Testing will evaluate detection quality, alert responsiveness, and system reliability. Add measured results after tests are performed; no performance figures are claimed here.

Planned Test Scenarios

Evaluation Metrics

Metric Purpose
Precision and recall Measure detection quality
False-alarm rate Measure unnecessary warnings
Missed-event rate Measure undetected risk events
Alert latency Measure time from detection to warning
Inference performance Measure edge-device processing performance
Network recovery Check synchronization after an outage
Access-control tests Verify role separation

Experiments are planned using suitable public datasets and a smaller, consent-based controlled dataset.

Detailed Budget

The preliminary hardware target is approximately LKR 30,000, assuming a Raspberry Pi can be borrowed. Final costs depend on component availability and hardware already accessible to the team.

Item Quantity Unit Cost Total
Raspberry Pi 1 Borrowed 0 LKR
Driver-facing camera 1 To be confirmed To be confirmed
ESP32 development board 1 To be confirmed To be confirmed
GPS module 1 To be confirmed To be confirmed
Buzzer, vibration motor, and LEDs 1 set To be confirmed To be confirmed
Contextual sensors and wiring 1 set To be confirmed To be confirmed
Power supply and enclosure 1 set To be confirmed To be confirmed
Preliminary target     30,000 LKR

This is a preliminary target, not a final cost breakdown. Replace estimates with actual quotations and quantities once confirmed.

Conclusion

This project aims to develop an affordable Edge-AI driver-vigilance and fleet-supervision prototype. By combining local computer vision, embedded alerts, contextual risk assessment, and a multi-user web application, the system is intended to help drivers respond to potential safety risks and enable fleet managers to review events across multiple vehicles.

Expected deliverables include a hardware prototype, local warning mechanism, backend and database, role-based dashboards, and an experimental evaluation report. Future work may include improving detection under challenging lighting conditions, optimizing edge inference, expanding fleet analytics, and pilot testing with transport operators.

Developed by Group 21 – Team Invictus as an Embedded Systems project at the University of Peradeniya.