FastKAN-DDD: A novel fast Kolmogorov-Arnold network-based approach for driver drowsiness detection optimized for...

Area of Science:

• Embedded Systems
• Machine Learning for Automotive Safety
• TinyML Applications

Background:

• Driver drowsiness is a major cause of traffic accidents, necessitating real-time fatigue detection systems.
• Existing machine learning (ML) and deep learning (DL) models are often too computationally intensive for resource-limited embedded systems.
• There is a need for lightweight, interpretable, and high-performance models for TinyML-based driver drowsiness detection in vehicles.

Purpose of the Study:

• To introduce FastKAN-DDD, a novel driver drowsiness detection model utilizing the Fast Kolmogorov-Arnold Network (FastKAN) architecture.
• To optimize the model for TinyML deployment through post-training quantization techniques.
• To evaluate the model's performance on a real-world driver drowsiness dataset for embedded automotive applications.

Main Methods:

• Developed FastKAN-DDD, incorporating learnable radial basis function (RBF)-based nonlinear activation functions for efficient function approximation.
• Applied post-training quantization (dynamic range, float-16, weight-only) to enhance suitability for TinyML deployment.
• Conducted experiments on the UTA-RLDD dataset, evaluating performance across different input resolutions and quantization schemes.

Main Results:

• Achieved a test accuracy of 99.94% on the UTA-RLDD dataset.
• Demonstrated extremely low inference latency (0.04 ms) and a minimal memory footprint (35 KB).
• Outperformed several state-of-the-art TinyML models in accuracy, computational efficiency, and model size.

Conclusions:

• FastKAN-DDD is exceptionally well-suited for real-time driver drowsiness detection on microcontroller-based systems.
• The model's lightweight and efficient design addresses the gap for TinyML applications in automotive safety.
• Publicly available code facilitates further research and development in this critical area.