Applicability of Neural Networks for Driving Style Classification and Maneuver Detection

Abstract

Maneuver and driving style detection are of ongoing interest for the extension of vehicle's functionalities. Existing machine learning approaches require extensive sensor data and demand for high computational power. For vehicle onboard implementation, poorly generalizing rule-based approaches are currently state of the art. Not being restricted to neither comprehensive environmental sensors like camera or radar, nor high computing power (both of what is today only present in upper class' vehicles), our approach allows for cross-vehicle use: In this work, the applicability of small artificial neural networks (ANN) as efficient detectors is tested using a prototypal vehicle implementation. During test drives, overtaking maneuvers have been detected 1.2 s prior to the competing rule-based approach in average, also greatly improving the detection performance. Regarding driving style recognition, ANN-based results are closer to targets and more patient at driving style transitions. A recognition rate of over 75 % is achieved.