A robust and adaptive driver identification framework for intelligent transportation systems

Abstract

Driver identification via behavioral characterization is a popular topic of intelligent transportation systems, yet existing methods often struggle with varying trip segmentation and vehicle type situations. This study proposes a refined deep learning framework to improve model robustness and adaptability under limited driving behavioral signal conditions. Our approach partitions driving trips into fixed-length identification windows, each with overlapping dynamic segments to allow the capture of temporal dependencies of driving patterns. Then, a deep neural network structure that combines a residual sequential autoencoder with an attention mechanism is incorporated to enhance the model identification performance through adaptive regularization. The framework is validated on two real-world datasets comprising 5 truck drivers and 5 sedan drivers through conventional train-validation-test. Our models achieve up to 91% accuracy for sedan drivers and 75% for truck drivers, significantly outperforming the baseline models. Notably, our approach maintains consistent performance across varying segment lengths, with an accuracy difference of only about 4% when the window length changes from 60 to 180 s. Experimental results demonstrate that our framework achieves strong segmentation-variability robustness and cross-vehicle adaptability.