Model Lightweighting for Real-time Distraction Detection on Resource-Limited Devices

Area of Science:

• Computer Vision
• Embedded Systems
• Artificial Intelligence

Background:

• Distracted driving poses a significant safety risk, necessitating accurate detection methods.
• Existing distracted driving detection systems often overlook the constraints of embedded devices, such as limited memory and computational power.
• There is a critical need for efficient algorithms capable of real-time performance on resource-constrained hardware.

Purpose of the Study:

• To develop a lightweight Convolutional Neural Network (CNN) for accurate and fast distracted driver detection on embedded systems.
• To optimize CNN architecture by adjusting block layers and reducing network channels without sacrificing detection accuracy.
• To enable real-time deployment of distracted driving detection models on various resource-limited devices.

Main Methods:

• Proposed a novel CNN light-weighting technique by modifying block layers and shrinking network channels.
• Developed a MobileNetV2-tiny model tailored for embedded applications.
• Deployed the optimized model on multiple devices for real-time performance evaluation.

Main Results:

• The MobileNetV2-tiny model achieved 1.63% higher accuracy on the AUC dataset compared to the original MobileNetV2.
• The proposed model utilizes only 78% of the parameters of the original MobileNetV2.
• Inference speed on resource-limited devices was, on average, 1.5 times faster than the original MobileNetV2, meeting real-time requirements.

Conclusions:

• The proposed CNN light-weighting method effectively enables accurate and fast distracted driver detection on embedded devices.
• The optimized model demonstrates superior performance in terms of accuracy, parameter efficiency, and inference speed.
• This approach addresses the underexplored problem of resource-limited distracted driving detection, paving the way for safer transportation.