A novel lane-changing decision model for autonomous vehicles based on deep autoencoder network and XGBoost

Abstract

Lane-changing (LC) is a critical task for autonomous driving, especially in complex dynamic environments. Numerous automatic LC algorithms have been proposed. This topic, however, has not been sufficiently addressed in existing on-road manoeuvre decision methods. Therefore, this paper presents a novel LC decision (LCD) model that gives autonomous vehicles the ability to make human-like decisions. This method combines a deep autoencoder (DAE) network with the XGBoost algorithm. First, a DAE is utilized to build a robust multivariate reconstruction model using time series data from multiple sensors; then, the reconstruction error of the DAE trained with normal data is analysed for LC identification (LCI) and training data extraction. Then, to address the multi-parametric and nonlinear problem of the autonomous LC decision-making process, an XGBoost algorithm with Bayesian parameter optimization is adopted. Meanwhile, to fully train our learning model with large-scale datasets, we proposed an online training strategy that updates the model parameters with data batches. The experimental results illustrate that the DAE-based LCI model is able to accurately identify the LC behaviour of vehicles. Furthermore, with the same input features, the proposed XGBoost-based LCD model achieves better performance than other popular approaches. Moreover, a simulation experiment is performed to verify the effectiveness of the decision model.