Driver status identification from driving behavior signals

Abstract

Driving behavior signals differ in how and under which conditions the driver uses vehicle control units, such as pedals, driving wheel, etc. In this study, we investigate how driving behavior signals differ among drivers and among different driving tasks. Statistically significant clues of these investigations are used to define driver and driving status models. Experimental results over the UYANIK database are presented. Driver identification over 23 drivers achieves a 57.39% identification rate with the fusion of gas and brake pedal pressure classifiers. Driver identification system with reduced number of drivers fits better on real-life scenarios. Driver identification rate within groups of three drivers is computed as 85.21%. Driver status identification over ten drivers with task and no-task classes yields a promising 79.13% task identification rate. Driving behavior is strongly related to past actions of drivers. In this study, we investigate driving behavior prediction from past driving signals. We propose a behavior prediction system, which performs temporal clustering of behavior signals and computes linear estimators for each temporal cluster. The temporal clustering is performed with hidden Markov model (HMM). Experimental evaluations show that distractive conditions have a certain effect on driving behavior, where the prediction errors are significantly increasing in these conditions. Road conditions are also influential on driving behavior prediction.