Driving Style Estimation for Impaired Driving Detection: An AI-based solution using Reinforcement Learning

Safety occupies a major spot when designing a new vehicle today. Carmakers give major importance to this aspect to make mobility and roads safer for vehicle passengers/occupants and pedestrians. The improvement of mobility safety is achieved through the optimisation of structural components, active systems, and the monitoring of driving conditions and driving style. The driving style is defined in terms of how the driver acts on the vehicle controls in response to external situations and stimuli. By characterizing this drive style, it is possible to identify psycho-somatic alterations, by identifying functional drifts with respect to the nominal driving modes. This work focuses on the latter and aims to develop an artificial intelligence-based system capable of detecting the behavioural patterns of a driver. Many Driver State Monitoring (DSM) methods are already present in vehicles, but they typically implement a camera-based solution that tracks eye movements or blink rate. However, European Union (EU) legislation suggests that the collection of biometric information should be avoided for privacy reasons. The aim of this project was therefore to develop a system that exploits a sensor-less architecture: no additional sensors are required in addition to those already present in the vehicle and no biometric measurements are needed. The driver's condition is estimated by means of an algorithm based on artificial intelligence that combines the driver's actions towards the vehicle (e.g., steering angle, accelerator/brake pedal pressure), vehicle dynamics (e.g., lateral/longitudinal acceleration) and, finally, the perception of the road and the vehicle's surroundings. The algorithm detects impaired driving conditions and issues an alarm. To realise the idea introduced above, the proposed methodology follows a rather innovative approach: the training of an autonomous agent with Reinforcement Learning in a virtual simulation environment that reproduces the road conditions and driving dynamics of a motor vehicle, which generates driver-specific ground-truth actions that feed a behavioural discriminator network capable of detecting impaired driving conditions as a functional drift from nominal conditions.