Cost-effective Driver Behaviour Detection System Using Deep Learning

Lawal, Comfort, Olaniyan, Olatayo M, Wejin, John S, Ogechukwu, Elekwa, Simonyan, Emmanuel O, Elias, Fanuel and Ekpo, Sunday C ORCID: https://orcid.org/0000-0001-9219-3759 (2024) Cost-effective Driver Behaviour Detection System Using Deep Learning. In: The Third International Adaptive and Sustainable Science, Engineering and Technology Conference (ASSET 2024), 16 July 2024 - 18 July 2024, Manchester, United Kingdom. (In Press)

Accepted Version File not available for download. Available under License In Copyright. Download (2MB)

Abstract

A World Health Organization (WHO) report highlights an average of 1.3 million lives lost annually to road accidents, with about 60% of these fatalities occurring in third-world countries with lower GDPs. Governments globally have implemented vehicle speed monitoring and traffic lights with vehicle plate recognition systems to meet Sustainable Development Goal (SDG) target 3.6. However, traditional speed monitoring systems often need help to track driver behaviour efficiently or are prohibitively expensive. Recent advancements in Machine Learning (ML) offer promising solutions for monitoring driver behaviour at a lower cost. This paper proposes a cost-effective method for detecting driver behaviour using deep learning models. A dataset of 13,669 images covering a range of poses and driver actions were collected using cameras from an 8MP Samsung Tab and an Apple Air iPad comprising images categorised into behaviours such as safe driving, texting, drowsiness, etc. Pre-processing steps include removing blurred images, annotating, and augmenting to address data imbalance and overfitting. Deep learning models, VGG16 and ResNet50, were fine-tuned to accept processed images. Evaluation results show 97% accuracy for VGG16 and 97.14% for ResNet50. This research underscores the potential of Artificial Intelligence (AI) and ML in developing cost-effective strategies to achieve SDG 3.6, particularly in low-income countries.