Driving behaviour monitor

Abstract

Driver behaviour monitoring and other Advanced Driver-Assistance Systems (ADAS) have received increasing attention in recent years for their contribution to improving road safety and reducing accidents. This research focuses on detecting manoeuvres considered dangerous for other vehicles such as abrupt braking, turning, and sudden lane changing. A sliding window approach was used, and various machine learning algorithms were tested, such as Gated Recurrent Neural Networks, Support Vector Machines and Random Forests. These were evaluated using driver behaviour datasets of real drivers together with data generated using the CARLA urban driving simulator. To generate a more comprehensive model for driver behaviour, the generated Carla dataset and the UAH-DriveSet were combined, achieving a 94% F1 score from a Recurrent Neural Network algorithm. Complex manoeuvres such as lane change events suffer from lower accuracy rates due to the limited window size and their similarity to other manoeuvres. A real-time driver behaviour evaluation application was developed, which implements an embedded model to perform real-time, on-device predictions. Inspired by the limited number of publicly available datasets in the field of driver behaviour monitoring, a field data collection application was developed. This application facilitates future research to develop various datasets through the user-defined, variable sampling rate, its large range of motion and position sensors, device permitting, and its output of pre-labelled data. Both applications achieve device independence through a rotation matrix which further boosts the usability of the application.