Personalized Driver Braking Behavior Modeling in the Car-Following Scenario: An Importance-Weight-Based Transfer Learning Approach

Zirui Li, Jianwei Gong, Chao Lu, Jinghang Li

Abstract: Accurately recognizing braking intensity levels (BIL) of drivers is important for guaranteeing the safety and avoiding traffic accidents in intelligent transportation systems. In this article, an instance-level transfer learning framework is proposed to recognize BIL for a new driver with insufficient driving data by combining the Gaussian mixture model (GMM) and the importance-weighted least-squares probabilistic classifier (IWLSPC). By considering the statistic distribution, GMM is applied to cluster the data of braking behaviors into three levels with different intensities. With the density ratio calculated by unconstrained least-squares importance fitting, the least-squares probabilistic classifier is modified as IWLSPC to transfer the knowledge from one driver to another and recognize BIL for a new driver with insufficient driving data. Comparative experiments with nontransfer methods indicate that the proposed framework obtains a higher accuracy in recognizing BIL in the car-following scenario, especially when sufficient data are not available.