A Comparative Study on Transferable Driver Behavior Learning Methods in the Lane-Changing Scenario

Abstract

The data-driven methods show many advantages in learning and recognizing driver behaviors and have been widely applied to build driver models. However, the data driven methods also raise a challenge for massive data collection and data processing since their performance usually relies on the quantity and coverage of training data. To reduce this reliance, transfer learning (TL) methods can be adopted. With the help of TL methods, the historical data of different drivers can also contribute to the new driver model and obtain a high accuracy, which can reduce data collection cost. In this paper, a new TL method semi-supervised balanced distribution adaptation (SS-BDA) is proposed based on distribution adaptation (DA). The proposed method can improve the performance of a new driver's decision-making model with insufficient data. Meanwhile, two major types of TL methods are compared and analyzed in building the personalized driver decision-making model in the lane change scenario. The comparative experiments are conducted using both the naturalistic data and simulated data with different TL methods and non-TL methods. Experimental results indicate that the proposed SS-BDA is capable of overcoming the model gap and achieves the best accuracy among all five methods. Besides, the TL method that adapts both marginal and conditional distribution performs better in driver model adaption.