A Driving Risk Prediction Approach Based on Generative Adversarial Networks and VANET for Autonomous Trams

Abstract

Driving safety is an essential prerequisite to the rapid development of autonomous trams. However, the relationship of driving risk factors is nonlinear, which makes modeling difficult. To improve the accuracy of driving risk prediction, a data driven approach based on Generative Adversarial Networks was proposed. First of all, a communication and alarming scenario of Vehicular Ad-hoc Networks was demonstrated, in which the original data sets can be collected and transmitted by the help of sensors and Road Side Units. Then the RFE feature selection algorithm was used to keep the key features. To deal the sample asymmetry problem, a DCGAN model was designed for sparse samples expansion. At last, the XGBoost algorithm was used to classification and output the risk prediction result. During the experiment implemented with the public and real data sets, the risk prediction accuracy of proposed approach can up to 97.24%, for which takes the advantages in generating of the sparse samples.