Blink Detection Using 3D Convolutional Neural Architectures and Analysis of Accumulated Frame Predictions.

Area of Science:

• Computer Vision
• Machine Learning
• Biomedical Engineering

Background:

• Blink detection is a valuable indicator for assessing clinical conditions and drowsiness.
• Accurate blink detection requires robust algorithms capable of analyzing video frame sequences.

Purpose of the Study:

• To propose and compare deep learning architectures for automated blink detection in videos.
• To evaluate the performance of different 3D convolutional neural network (CNN) models and a 3D autoencoder for blink detection.

Main Methods:

• Eye regions were extracted using an eye detector and formatted as 3D input (300 ms temporal span).
• Two 3D CNNs (simple 3D CNN, 3D ResNet) and a 3D autoencoder with a latent space classifier were implemented.
• A frame prediction accumulator combined with morphological processing and watershed segmentation was used for blink detection and temporal localization.

Main Results:

• The proposed framework, particularly the 3D ResNet coupled with the prediction accumulator, achieved favorable results compared to state-of-the-art methods.
• The 3D ResNet demonstrated superior performance and speed in blink detection and temporal localization.
• The system was trained on data from 10 participants and tested on 5, analyzing 162,400 frames and 1172 blinks per eye.

Conclusions:

• Deep learning architectures, especially 3D ResNet, are highly effective for automated blink detection in video.
• The proposed prediction accumulator enhances blink detection accuracy and temporal resolution.
• This approach holds significant potential for real-time drowsiness monitoring and clinical applications.