AcademicAI & Machine Learning•2023
Driver Drowsiness Detection
Computer vision system that monitors driver eye closure, head pose, and facial landmarks to detect drowsiness and trigger alerts. Designed for automotive and fleet management applications.
Project Overview
Computer vision system that monitors driver eye closure, head pose, and facial landmarks to detect drowsiness and trigger alerts. Designed for automotive and fleet management applications.
Problem Statement
Driver fatigue is a leading cause of road accidents. Commercial fleets need reliable, non-intrusive monitoring systems that work in real-world conditions (varying lighting, glasses, facial hair).
Proposed Solution
Multi-modal drowsiness detection combining eye aspect ratio (EAR), head pose estimation, and facial action units, with adaptive thresholds and minimal false positives.
Key Features
Eye aspect ratio (EAR) and PERCLOS calculation
Head pose estimation (yaw, pitch, roll)
Facial landmark detection with MediaPipe/dlib
Adaptive thresholding for individual baselines
Multi-stage alerting (visual, auditory, haptic)
Fleet management dashboard with trip reports
Night/low-light performance with IR camera support
Glasses and sunglasses compatibility
Edge deployment on Raspberry Pi / Jetson Nano
GDPR-compliant data handling
Technology Stack
PythonOpenCVdlibMediaPipeNumPyArduino/Raspberry PiReactTypeScriptNode.jsInfluxDBGrafanaDocker
Project Screenshots
Development Process
01
Algorithm Development
4 weeks- EAR/PERCLOS implementation
- Head pose estimation pipeline
- Sensor fusion algorithm design
02
Embedded Optimization
3 weeks- Model quantization
- Pi/Jetson deployment
- Thermal and power profiling
03
System Integration
3 weeks- Hardware integration (camera, buzzer, haptic)
- Dashboard development
- Alert escalation logic
04
Testing
2 weeks- Simulator testing with 20 drivers
- Real-world fleet pilot (5 vehicles)
- Safety validation
Challenges
- Robust detection in varying lighting conditions
- Minimizing false positives during normal driving
- Real-time performance on embedded hardware
- Individual calibration without explicit setup
Results & Outcomes
- 92% drowsiness detection accuracy in controlled tests
- <5% false positive rate during alert driving
- 30 FPS on Raspberry Pi 4 with optimized model
- Integrated with fleet telematics platform