Driver Identification Through Multi-State Car Following Modeling

Driver profiling and characterization is one of the key issues that models heterogeneities inter- and/or intra-drivers' behaviors, which forms the basis of many applications such as driver identification, abnormal driving state awareness, personalized driving assistance, etc. This research proposes a driver identification method through multi-state car following modeling, where a GMM (Gaussian Mixture Model) is used to model the internal stochasticity of an individual driver using a mixture of Gaussian kernels, a driver profile is defined to describe the probabilities of a car following behavior being operated by each registered driver, a projection matrix is learned to project a feature vector to a lower dimensional space that has the best discriminative nature for GMM modeling, driver profiling and ID prediction. Experiments are conducted using the naturalistic driving data that are collected on the motorways in Beijing, where a dataset of eight-drivers' car-following data is generated. Promising results are demonstrated and the performance of driver identification on different aspects are examined.