Semi-Supervised Estimation of Driving Behaviors Using Robust Time-Contrastive Learning

Estimation of driving behaviors is an elemental technology in a driving support system for a vehicle. For realizing intelligent estimation of driver behaviors, many studies have explored the use of machine learning methods mainly in a supervised fashion that require a large amount of labeled driving data. In this study, we hypothesize that the time-contrastive learning (TCL) could be helpful for reducing the number of labeled data for the supervised learning and numerically tested it using a public data set. For this purpose, we constructed three models to estimate driving behaviors from vehicle dynamics: 1) a naive linear classifier implemented by linear discriminant analysis (LDA) model; 2) an LDA classifier combined with a feature extraction process by the original TCL; 3) the same as 2) except the robust version of TCL was employed instead of the original TCL. The results were not supportive to our hypothesis: Model 1) showed better performance than the other models when very few labeled data was available; and two models with TCL outperformed the other without TCL for a considerable number of labeled data. We conclude discussions on some limitations of this study and open issues for the future.