Comparative Study of VGG16, ResNet50, and YOLOv8 Models in Detecting Driver Distraction in Varying Lighting Conditions

Observing driver distractions while driving gives valuable information to prevent accidents, so it is necessary to use effective monitoring methods. Deep learning is showing new capabilities in solving this issue. This study evaluates the results of CNN, YOLOv8, ResNet50 and VGG16 deep learning models as they detect drivers who are practising distracted driving behaviours under real-time and various lighting conditions (day and night). The models were trained on two datasets: the labelled State Farm dataset and the Driver Monitor Dataset (DMD). They successfully identified ten distinct categories of distraction for the State Farm dataset and five categories for the monitoring drivers dataset. Pre-trained models were optimized using transfer learning through fine-tuning to enhance detection accuracy. This paper studies related work on distracted driving and shares ideas for designing advanced systems that use various methods to improve accuracy. YOLOv8 reached an outstanding test accuracy of 98.46% on the State Farm dataset, proving itself superior to other methods and confirming its effectiveness for monitoring. In addition, YOLOv8 reached 96.46% accuracy in the DMD dataset, outperforming VGG16 at 90.58% and ResNet50 at 70.80%. YOLOv8 was able to recognise important driver behaviours in real time with a dataset of 15 subjects and 20 different driving postures. The research proves that the YOLOv8 model is fit for use in intelligent monitoring systems designed to detect distracted driving and promote safer driving through focused actions.