Categorization of driving characteristics using deep clustering

Abstract

Driving behavior is an effective means to analyze a driver who drives a vehicle. In order to reduce traffic crashes and environmental pollution problems caused by the increase in vehicles, various studies on the driver are being conducted. In particular, due to the recent development of various sensor technologies, it is possible to collect large-scale driving records from many drivers. However, the large-scale driving record is multidimensional time-series data, and there is a limit to quantitatively analyzing it. The purpose of this dissertation is to categorize driving characteristics from driving records and conduct driving behavior analyses using the characteristics. A methodology for deriving the elementary driving behavior (EDB), which is a norm driving behavior for each driving environment, is presented through deep clustering, and the EDB is extracted using the driving record data of taxi drivers. For a traffic safety study, the EDB-based abnormal driving score that can numerically represent drivers' propensities for each driving environment is developed, and the differences according to driver propensities are verified. For a traffic environment study, the EDB-based driving cycle generation model is developed to reflect various trajectories' characteristics. The driving behavior characterization method presented in this dissertation can be effectively used in transportation research in the mobility era, contributing to the construction of a safe and efficient transportation system.