Driver and Vehicle Unsafe Behavior Tracking using Deep Learning

Abstract

Addressing the global concern of road safety requires focused attention, with driver behavior playing a crucial role in contributing to road accidents on a global scale. Establishing a reliable methodology that takes into account both driver and car behavior is essential for accurate assessment and monitoring. This enables responsible institutions, such as governments, insurance companies, and fleet-owning businesses, to effectively oversee and regulate driver conduct on the road, thereby enhancing overall road safety. This research explores seven driver behaviors and three vehicle behaviors. Logistic Regression achieved 89% accuracy in detecting phone calls, while Fine-tuned YOLO showed a notable 99% mAP50 for hands-off-steering-wheel instances. Fine-tuned YOLO also achieved an overall mAP50 of 97% for behaviors like drinking, eating, smoking, and mobile use. In car behavior classification, CNN-LSTM led with a macro f1 score of 71.22%, while 1D-CNN excelled with a micro f1-score of 86.78%, emphasizing the importance of advanced models in comprehensive road safety assessments.