DDD TinyML: A TinyML-Based Driver Drowsiness Detection Model Using Deep Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Embedded Systems

Background:

• Driver drowsiness is a significant factor in traffic accidents.
• Integrating deep learning (DL) models with Internet-of-things (IoT) devices for driver drowsiness detection is challenging due to limited IoT resources.
• Real-time applications require low latency and lightweight computation, which are difficult to achieve with traditional DL models.

Purpose of the Study:

• To investigate the application of Tiny Machine Learning (TinyML) for driver drowsiness detection on microcontrollers.
• To evaluate and compare the performance of lightweight DL models optimized for size and accuracy.
• To assess the effectiveness of different quantization methods in reducing model size and improving accuracy.

Main Methods:

• An overview of TinyML principles and methodologies was presented.
• Five lightweight DL models (SqueezeNet, AlexNet, CNN, MobileNet-V2, MobileNet-V3) were proposed and evaluated.
• Quantization techniques, including quantization-aware training (QAT), full-integer quantization (FIQ), and dynamic range quantization (DRQ), were applied to optimize models.

Main Results:

• The CNN model achieved the smallest size (0.05 MB) using DRQ.
• MobileNet-V2 achieved the highest accuracy (0.9964) after DRQ optimization.
• SqueezeNet and AlexNet also demonstrated high accuracy (0.9951 and 0.9924, respectively) using DRQ.

Conclusions:

• TinyML is a viable approach for deploying efficient driver drowsiness detection systems on microcontrollers.
• Optimized lightweight DL models, particularly MobileNet-V2 with DRQ, offer a promising solution for real-time, resource-constrained applications.
• The study highlights the potential of TinyML to enhance road safety by enabling on-device intelligent systems.