Pupil Detection Algorithm Based on ViM.

Area of Science:

• Computer Vision
• Machine Learning
• Biomedical Engineering

Background:

• Pupil detection is crucial for human-computer interaction, driver monitoring, and medical diagnostics.
• Current algorithms struggle with variable lighting and occlusions, limiting their real-world application.
• Robust and efficient pupil detection remains a significant challenge.

Purpose of the Study:

• To propose a novel pupil detection algorithm, ViMSA, that addresses limitations of existing methods.
• To enhance the accuracy, robustness, and efficiency of pupil detection.
• To demonstrate ViMSA's generalization capability across diverse datasets and conditions.

Main Methods:

• Developed ViMSA algorithm based on the ViM model, incorporating weighted feature fusion.
• Integrated multi-head self-attention (MSA) for global feature integration.
• Utilized Fast Fourier Transform (FFT) to optimize MSA computational complexity.

Main Results:

• Achieved 99.6% detection accuracy with an RMSE of 1.67 pixels.
• Exceeded 100 FPS processing speed, meeting real-time requirements.
• Demonstrated exceptional generalization across 30 datasets with ~135,000 images.

Conclusions:

• ViMSA offers a significant advancement in pupil detection technology.
• The algorithm is robust under variable lighting and occlusion, suitable for real-time applications.
• ViMSA has broad applicability in automotive safety, assistive technology, and human-computer interaction.