Enhancing Driving Safety: Driver Behavior Analysis Via Lightweight CNN Model

Aggressive driving poses a significant global challenge, contributing to numerous traffic accidents and jeopardizing road safety. While considerable research has explored predictive methods for such behaviors using diverse data and machine learning approaches, there has been limited focus on lightweight convolutional neural networks (LCNNs) for analyzing driver behavior. This study aims to address this gap by introducing an LCNN-based framework to enhance driver behavior analysis. The research begins with dataset acquisition from Kaggle, followed by the development of a novel preprocessing pipeline, incorporating label encoding and standardization techniques to optimize the data for analysis. The study explores driver behavior prediction through experiments utilizing 1D-CNN and 2D-CNN, called LCNN models. A robust evaluation approach was adopted, employing 100-repetition holdout validation. For cross-validation, configurations of 6×1 for 1D-CNN and 3×2×1 for 2D-CNN were implemented. Performance metrics, including training and testing accuracy, mean squared error (MSE), and area under the curve (AUC), were analyzed to compare the models. Results revealed that the 1D-CNN model achieved 100% accuracy, while the 2D-CNN model demonstrated a slightly lower accuracy of 99.74%. This research underscores the effectiveness of GPS data in predicting aggressive driving behavior, providing crucial insights to enhance road safety initiatives.