Drowsy driver detection system using eye blink patterns

Taner Danışman, Ioan Marius Bilasco, Chabane Djeraba, Nacim Ihaddadene2010 · DOI

TL;DR

This paper introduces a low-cost, real-time system for detecting drowsy drivers by analyzing eye blink patterns using a standard webcam. The novel approach leverages horizontal eye symmetry to detect blinks, achieving high accuracy and speed without requiring specialized hardware or complex initialization.

What Problem Does This Paper Address?

Drowsy driving is a significant contributor to fatal traffic accidents, often cited as the primary factor in single-vehicle run-off-road crashes. While major automotive companies have developed drowsy driver detection systems (DDDS), these often rely on expensive hardware like infrared cameras and may lack full automation. This paper aims to address this gap by proposing an affordable and fully automatic DDDS that can detect drowsiness by monitoring eye blink duration and patterns, thus potentially preventing accidents.

Methods

The proposed system comprises several key stages:

  • Face Detection: Utilizes the Viola-Jones face detector, a well-established algorithm for real-time face recognition.
  • Eye Detection: Employs a neural network-based eye detector from the STASM library to accurately locate pupil positions.
  • Face Orientation Correction: To mitigate the effects of head tilting, the system estimates face orientation using the vertical positions of both eyes and corrects it by rotating the frame.
  • Eye Blink Detection: This is the core novelty. The system extracts a small region around the pupil and applies contrast stretching. It then divides this region into upper and lower halves. The key insight is that an open eye exhibits horizontal symmetry, while a closed eye does not. This symmetry property is quantified using a difference image and a cumulative sum to determine if the eye is open or closed. This method works on a single frame, eliminating the need for sequential frame analysis for initial detection.
  • Drowsiness Detection: The system categorizes drowsiness into three levels (Awake, Drowsy, Sleeping) based on predefined thresholds for eye blink durations (TDrowsy = 400ms, TSleeping = 800ms).

Datasets

The primary dataset used for evaluation is the JZU Eyeblink database. This database includes videos of users, some wearing glasses, under various conditions (static/moving faces, dynamic backgrounds, different lighting). The authors also mention manual annotation of frames from this database to establish ground truth for eye blinks. They note that no common database exists for comparing drowsiness detection results directly, and they plan to work with the Av@Car database for future comparisons.

Key Insights and Ideas

  • Horizontal Eye Symmetry as a Blink Indicator: The paper's central innovation is using the horizontal symmetry of the eye region to distinguish between open and closed states. This is a clever and computationally efficient approach that avoids complex feature extraction or temporal modeling for initial blink detection.
  • Single-Frame Detection Capability: The ability to detect eye state from a single frame is a significant advantage for real-time applications, reducing computational load and complexity compared to methods requiring multiple frames.
  • Low-Cost, High-Performance Solution: By relying on a standard webcam and efficient algorithms, the system offers a cost-effective alternative to existing DDDS solutions that often require specialized hardware.
  • Robustness to Head Movements: The face orientation correction step enhances the system's robustness to common head movements, a limitation in some previous approaches.

Key Results

  • Accuracy: The proposed system achieves a 94.8% accuracy in detecting eye blinks on the JZU database.
  • Precision and Recall: Precision is reported at 90.7%, and Recall at 71.4%.
  • False Positive Rate (FPR): A very low 1% FPR is achieved.
  • Speed: The system runs at 110 frames per second (fps) on a 2.9 GHz Intel Xeon CPU with a 320x240 resolution, making it suitable for real-time operation.

Limitations and Future Work

  • Presence of Glasses: The authors acknowledge that the presence of glasses significantly affects the performance of core components like face and eye detection, and symmetry calculation.
  • High Illumination Changes: Extreme lighting variations can also impact the system's accuracy.
  • Lack of Standard Drowsiness Database: The absence of a common benchmark for drowsiness detection makes direct comparison with other drowsiness detection methods challenging.
  • Future Work: The authors plan to investigate the Av@Car database for further comparisons and to address the limitations posed by glasses and lighting conditions.

Summary updated

Abstract

International audience
BibTeX
@misc{dansman2010drowsy,
  author = {Taner Danışman and Ioan Marius Bilasco and Chabane Djeraba and Nacim Ihaddadene},
  title = {{Drowsy driver detection system using eye blink patterns}},
  year = {2010},
  doi = {10.1109/icmwi.2010.5648121},
}
Figure