Enhancing Driver Distraction Detection Through the Synergy of Deep and Traditional Machine Learning

Distracted driving is a major contributor to motor vehicle accidents, causing injury and loss of life. It is one of the major factors that affect the overall driving behavior of a person. Insurance companies take into consideration factors like gender, age, etc. to set insurance premiums for their customers. Today, machine learning and artificial intelligence can eradicate this bias. A machine learning model can analyze driving behavior, such as the frequency and severity of accidents, the speed at which they drive, and their habits such as distracted driving. Based on this information, the model can then determine the risk of accidents for the driver and set a corresponding insurance premium. In this project, we initiated multiclass driver distraction classification through transfer learning, utilizing pre-trained models such as ResNet50 and VGG16 to achieve 92.2% and 94.1% accuracies. Subsequently, we introduced a custom CNN model which attained a validation accuracy of 98.7%. To enhance practicality and accuracy, we transformed the multiclass task into binary classification. Here, we leveraged the fusion of deep learning and traditional machine learning, treating the convolutional filters as a feature extractor and employing Random Forest for binary classification, resulting in an impressive 99.6% validation accuracy. Our model was trained on the State Farm Distraction dataset that is available to the public on Kaggle. Additionally, we computed a driver-specific distraction score, offering potential applications in assessing accident risk, setting equitable insurance premiums, and promoting safer driving habits