Driver Identification Based on Stop-and-Go Events Using Naturalistic Driving Data

There are many occasions on which several drivers share a vehicle today. In order to provide better personalized service, Advanced Driver Assistance Systems (ADAS) needs to identify who amongst a small set of potential drivers is currently behind the wheel. In this paper, we present a new approach for driver identification based on Stop-and-Go events which are recognized and extracted from the naturalistic driving data. Random Forests algorithm is chosen to filter important features and built model. In the research, Stop-and-Go events are divided into stop, waiting, and go phases. Controller Area Network (CAN) and three-axis acceleration signals during the three phases including vehicle speed, longitudinal acceleration, brake pedal position and engine revolutions per minute are used to characterize driver behaviors. The driving behaviors show a difference at each phase of Stop-and-Go events. In the experiments, 553 Stop-and-Go events of 6 drivers among 55 drivers are selected to analyze concretely. Experiments show the Random Forests model covering all phases has the best performance to identify drivers. Moreover, the effect of different data sources including speed with acceleration, engine revolutions and brake pedal position on the model performance were discussed. The model using all data sources performs best with average accuracy 91.2% in 10-fold cross validation. The proposed model using Stop-and-Go events demonstrates comparable performance with the existing state-of-the-art studies. Especially, the baseline model can reach 100% prediction accuracy by using voting strategy. The baseline model requires only speed data which makes it much easier to be put into use in industry.