Learning-based probabilistic modeling and verifying driver behavior using MDP

Assisted driving has always been a hot research issue. The existing work mainly focuses on modeling vehicles behavior. However, there still lacks research work of modeling and verifying driver behavior. To solve these problems, we are committed to modeling and analyzing the driver behavior with Markov Decision Process (MDP). The aim is to achieve safe driving by monitoring and predicting the driver's states. In this paper, we propose a novel approach to construct MDP models of driver behavior. It comprises four phases: (1) data preprocessing using Convolutional Neural Network (CNN), wherein we adopt CNN to extract the features of driver behavior with the simulation data; (2) Bayes-based learning, wherein we construct a training set and use the Naive Bayes algorithm to train the State Prediction Model (SPM); (3) MDP generating, wherein we propose an algorithm to generate MDP models for the driver behavior with the help of SPM; and (4) quantitative analysis, wherein we analyze the uncertain behavior of the driver with probabilistic model checking technology. The main novelty of our work is to model and verify the driver behavior by integrating the learning and the model checking technology. To implement our approach, we have developed the MDP generator. Moreover, the quantitative analyses of the driver behavior are conducted with the model checker PRISM. The experiment results show that our approach facilitates generating MDP models, which helps to model and analyze the uncertain behavior of the driver.