Drowsiness Detection System

This project implements a drowsiness detection system using multiple face detection models and facial landmark analysis. It leverages computer vision techniques to detect drowsiness based on Eye Aspect Ratio (EAR) and mouth opening distance.

Prerequisites

• Python 3.6+
• Google Colab/Jupyter Notebook (recommended for GPU support)
• Basic understanding of computer vision concepts

Installation

# Clone YOLOv5 repository
!git clone https://github.com/ultralytics/yolov5.git
!pip install -r yolov5/requirements.txt

# Install dependencies
!pip install dlib imutils opencv-python-headless numpy retinaface_pytorch pydub

# Download facial landmark predictor
!wget http://dlib.net/files/shape_predictor_68_face_landmarks.dat.bz2
!bunzip2 shape_predictor_68_face_landmarks.dat.bz2

# Download Caffe model files (SSD Face Detector)
!wget https://raw.githubusercontent.com/opencv/opencv/master/samples/dnn/face_detector/deploy.prototxt
!wget https://raw.githubusercontent.com/opencv/opencv_3rdparty/dnn_samples_face_detector_20180205_fp16/res10_300x300_ssd_iter_140000_fp16.caffemodel

Model Approaches

1. Caffe SSD Face Detector with dlib Landmarks

Files Required:

• deploy.prototxt
• res10_300x300_ssd_iter_140000_fp16.caffemodel
• shape_predictor_68_face_landmarks.dat

Steps:

1. Face detection using Caffe-based SSD
2. Landmark detection with dlib's 68-point predictor
3. EAR calculation for both eyes
4. Drowsiness determination (EAR threshold: 0.3)

2. YOLOv5 Face Detector with Facial Landmarks

Files Required:

• shape_predictor_68_face_landmarks.dat
• YOLOv5s pretrained weights (automatically downloaded)

Steps:

1. Face detection using YOLOv5
2. Landmark detection with dlib
3. EAR calculation + mouth opening detection
4. Combined drowsiness analysis (EAR threshold: 0.279)

3. RetinaFace with dlib Landmarks

Files Required:

• shape_predictor_68_face_landmarks.dat
• RetinaFace ResNet50 weights (automatically downloaded)

Steps:

1. Face detection using RetinaFace
2. Landmark detection with dlib
3. EAR-based drowsiness detection (threshold: 0.25)

Project Structure

├── yolov5/                   # YOLOv5 implementation
├── Video1.mp4                # Input video file
├── shape_predictor_68_face_landmarks.dat  # Landmark model
├── deploy.prototxt           # Caffe model configuration
├── res10_300x300_ssd_iter_140000_fp16.caffemodel  # Caffe weights
└── Drowsiness_Detection.ipynb  # Main implementation notebook

Usage

1. Video Input Preparation:

   • Place your video file in the project directory
   • Update video_path in code to match your filename

2. Running Detection:

# For Caffe-based detection
main()

# For YOLOv5-based detection
model = torch.hub.load('ultralytics/yolov5', 'yolov5s', pretrained=True)
processed_image = detect_drowsiness_and_draw(image, model, predictor_path)

# For RetinaFace detection
detect_faces(video_path)

Key Features

• Multi-model face detection support
• Real-time EAR calculation
• Mouth opening detection
• Visual feedback with bounding boxes and landmarks
• Configurable thresholds for different scenarios

Results

• DROWSY state indicated by red text when EAR < threshold
• NOT DROWSY state shown in green when eyes are open
• Bounding boxes around detected faces
• White landmarks highlighting eyes and mouth regions

Threshold Tuning

Model	Default EAR Threshold	Mouth Open Threshold
Caffe + dlib	0.30	N/A
YOLOv5 + dlib	0.279	15 pixels
RetinaFace + dlib	0.25	N/A

Adjust these values in code based on your specific requirements.

Acknowledgements

• YOLOv5 by Ultralytics
• dlib's facial landmark predictor
• OpenCV's Caffe models
• RetinaFace PyTorch implementation