An Efficient Distracted Driving Detection Based on MobileNet V2SE Fusion

Abstract

The issue of distracted driving has become a significant concern, leading to numerous fatalities and injuries. There is a pressing need to develop innovative approaches to identify and mitigate this problem. This paper proposes a lightweight deep learning model that uses MobileNetV2 as the base and includes attention mechanisms like the Squeeze and Excite (SE) module to identify distracted driver actions. The proposed model underwent rigorous training and testing using the American University in Cairo (AUC) distracted driver dataset, which includes ten distraction categories. The model was optimized through hyperparameter tuning, data augmentation, and class weighting. To validate the model’s effectiveness, a confusion matrix, frames per second (FPS), accuracy, precision, recall, and F1 score were used as evaluation metrics. The proposed model achieved 93% accuracy with a batch size of 32, learning rate of 0.0001, and 21 epochs. Furthermore, the proposed model was compared to the MobileNetV2 and other existing architectures regarding accuracy and parameters. The proposed method outperformed unmodified deep learning models and maintained a balance between accuracy and parameter utilization, while some other modified models performed slightly better. The proposed method shows promising potential for accurately detecting distracted drivers with efficiency.