Deep Neural Networks for Driver Identification Using Accelerometer Signals from Smartphones

Abstract

With the evolution of the onboard communications services and the applications of ride-sharing, there is a growing need to identify the driver. This identification, within a given driver set, helps in tasks of antitheft, autonomous driving, fleet management systems or automobile insurance. The object of this paper is to identify a driver in the least invasive way possible, using the smartphone that the driver carries inside the vehicle in a free position, and using the minimum number of sensors, only with the tri-axial accelerometer signals from the smartphone. For this purpose, different Deep Neural Networks have been tested, such as the ResNet-50 model and Recurrent Neural Networks. For the training, temporal signals of the accelerometers have been transformed as images. The accuracies obtained have been 69.92% and 90.31% at top-1 and top-5 driver level respectively, for a group of 25 drivers. These results outperform works in the state of the art, which can even utilize more signals (like GPS- Global Positioning System- measurement data) or extra-equipment (like the Controller Area-Network of the vehicle).