Deep Learning Prediction of Vehicle Lane Departure During Night-Times: A Synthetic Over-Sampling Framework with Enhanced Dimensionality Reduction

The inevitability and imminent widespread adoption of autonomous vehicles are acknowledged, yet human supervision in driving remains crucial for the foreseeable future. Assessing a driver's capacity to take control during critical situations involves monitoring lane departure events during night-times through the incorporation of various features. In response to this challenge, this study endeavors to develop a deep learning framework designed to predict lane departure events based on driver demographic traits including age and gender, further to driver in- puts like throttle and brake pedal positions, along with vehicle data such as speed and tires’ temperature obtained through a driving simulator. Dimensionality reduction has been endorsed using Random Forest and Principal Component Analysis to enhance the system’s performance. Moreover, the SMOTE-TL and ADASYN synthetic oversampling techniques were employed to resolve the data imbalance issue. The most favorable outcomes were achieved through the proposed deep learning MLP model. Notably, there has been limited exploration into assessing vehicle lane departure events during night times based on the proposed framework. In its entirety, the proposed system and its findings offer novel insights and present a promising tool for enhancing road safety within the realm of traffic intelligent transportation systems.