Design and Implementation of a Hybrid Fuzzy-Reinforcement Learning Algorithm for Driver Drowsiness Detection Using a Driving Simulator

Driver drowsiness is the cause of many fatal accidents all over the world. Many research works have been conducted on detecting driver drowsiness for more than half a century, but statistical data show that such accidents have not decreased significantly. Most researchers have focused on using certain sensors and extracting their relevant features. However, there has been no research work on developing an algorithm to detect driver drowsiness independently from the input type. In this paper, a hybrid fuzzy-reinforcement learning drowsiness detection algorithm is presented. This algorithm is flexible to work with any number and any kind of data related to driver alertness. It estimates the level of alertness based on an arbitrary number of inputs. The algorithm extracts driving patterns specific to each driver and determines driver’s level of drowsiness using a continuous numerical variable rather than a discrete variable. To evaluate the algorithm, only six features related to only steering wheel angle and velocity are used. The accuracy of the user-specific data is 81.1% validated with the Observer Rating of Drowsiness criterion. This hybrid fuzzy-reinforcement learning algorithm has 46.4% improvement over the artificial neural network user-specific dataset method, and 49.2% over the artificial neural network general dataset method. The results can be improved even further if we use more features related to the driver and the vehicle.