Spatiotemporal CNN with Pyramid Bottleneck Blocks: Application to eye blinking detection

Area of Science:

• Computer Vision
• Biometrics

Background:

• Eye blink detection is crucial for facial analysis tasks like anti-spoofing and driver drowsiness detection.
• Existing methods often fail in uncontrolled environments ('in the wild') and struggle with multiple blinks in a sequence.

Purpose of the Study:

• To develop a fast and robust framework for detecting and verifying multiple eye blinks from image sequences.
• To address challenges like varying lighting, poses, and appearance changes in real-world scenarios.

Main Methods:

• Utilizes a fast facial landmark detector to identify key facial points, including eye regions.
• Employs a Singular Value Decomposition (SVD)-based method within a sliding window for initial blink detection.
• Verifies detected blink candidates using a 2D Pyramidal Bottleneck Block Network (PBBN).
• An alternative approach uses a continuous 3D PBBN for frame sequences.

Main Results:

• The proposed framework effectively extracts multiple blinks from image sequences under challenging conditions.
• Experimental results demonstrate superior performance compared to existing state-of-the-art approaches.

Conclusions:

• The developed framework offers a significant advancement in eye blink detection, particularly for real-world applications.
• The approach shows promise for enhancing the accuracy and reliability of facial analysis systems.