Driver recognition and driver verification using data mining techniques

In this thesis we present our research in driver recognition and driver verification. The goal of this study is to investigate the affect of different classifier fusion techniques on the performance of driver recognition and driver verification. We are using five different driving behavior signals for identifying the driver identities. Driving features were extracted from these signals and Gaussian Mixture Models were used for modeling the driver behavior. Gaussian Mixture Model training was performed using the well-known EM algorithm. In recognition study posterior probabilities of identities called scores were obtained with the given test data. These scores were combined using different fixed and trainable (adaptive) combination methods. In verification study we compared posterior probabilities with fixed threshold values for each classifier. For different thresholds, false-accept rate versus falsereject rate was plotted using the receiver operating characteristics curve. We observed lower error rates when we used trainable combiners. We conclude that combined multi-modal signal or classifier methods are very successful in biometric recognition and verification of a person in a car environment.