Driving Style Estimation for Driver State Monitoring Truthful Data Acquisition via Driving Simulator

Today's studies on Driving Style Estimation (DSE) provide valuable information on the subjectivity of different drivers behind the wheel, which directly impacts road safety, vehicle usability and component wear. The variety of use cases where this area of study is applied include impaired driving detection (e.g. distraction, drowsiness, aggressiveness), driver attitude profiling (e.g. theft situation, insurance), mission optimization (e.g. fuel reduction, battery wear reduction in case of electric vehicles), among many others. Due to the high cost of obtaining real driving data, the adoption of driving simulators has been the best alternative, even for big car manufacturers, to acquire realistic driving behavior data. Such data gets processed to build an accurate driver model, which is considered a valuable asset to inform state-of-the-art Advanced Driver Assistance Systems (ADAS) about human driving subjectivity. In this study, a pre-existing MATLAB/Python driving simulator framework is taken and improved to develop a robust software in order to obtain more accurate results for impaired driving detection. Furthermore, a physical driving simulator setup is mounted for human-in-the-loop testing with the aim of high-quality driving data acquisition and processing, in addition to the development of a Graphical User Interface (GUI) and realistic scenarios exploiting Simulink's Unreal Engine 4 API. Furthermore, exploiting the driving simulator tool allows for naturalistic driving data collection processes which follow international standards and render the neural network learning process more efficient, giving the different algorithms visibility to realistic driving data and behavior. Ultimately, several improvements on each stage of the process, among which model calibration and hyper-parameter tuning, yield a more stable and better behaving impaired detection system, exploiting the full capability of data-driven methods for safety critical systems, calling for a bigger participation on today's automotive systems and a call to merge with traditional model-based methods.