Driving Behavior Modeling and Characteristic Learning for Human-like Decision-Making in Highway

To make autonomous vehicles consider driver's personalized characteristics, this paper proposes an integrated model and learning combined (IMLC) algorithm to realize human-like driving. It includes the integrated driving behavior modeling to ensure basic safety and the characteristic learning to further imitate human driver's style. Firstly, an integrated behavior model is built according to driver's operation logics, including lane advantage assessment, target lane selection and acceleration determination. The lane advantage is assessed by five lane features, like safety, efficiency, cooperativity, etc. Then, parameters of the built model are learned from human's demonstrations. For the lane selection parameter, a novel lane feature extraction method is presented and the maximum entropy inverse reinforcement learning (IRL) is adopted to solve. For the acceleration parameter, since it's hard to extract human's acceleration features accurately, the particle filtering is used to estimate. Finally, the IMLC algorithm is validated in highD dataset compared to existing algorithms. The results show that the RMSE of position and velocity in 9s are within 4.2 m and 0.9 m/s, which has great advantage. Moreover, we test the human-like performance in driver simulator. The safety and efficiency in this process are fairly approximate.

Automated summary

This paper proposes an integrated model and learning combined (IMLC) algorithm to achieve human-like driving for autonomous vehicles. The algorithm includes integrated driving behavior modeling and characteristic learning. The algorithm is validated using highD dataset, and the results show that it has great advantages in position and velocity accuracy.