Lightweight YOLO for distracted driver detection on edge devices

Abstract

Edge AI, with its ability to process data locally on devices within vehicles, presents a promising approach to real-time driver monitoring. However, despite advancements in robust deep learning-based distracted driver detection, there is a critical gap in research on deploying these methods on edge devices. Real-world applications demand a balance between accuracy and real-time inference speed on resource-constrained devices. This work addresses this challenge by investigating the performance of a lightweight, human activity recognition-based distracted driver detection method. A comparative analysis study is conducted to compare the performance of four lightweight YOLO models. The study also explores the generalisability of the approach for driver distraction detection across four public datasets. Experimental results reveal that the tiny version of the YOLOv7 object detector provides the best balance between accuracy and inference speed. The algorithm achieved an average F1-score of 0.45 across four datasets and an average inference speed of 21.97 ms or 46 frames per second.