Toward intuitive drift assist control: driver drift intention recognition using a data-based approach

Abstract

In contrast to autonomous drifting where path-planning determines when and where the vehicle drifts, to support drift-assist control systems in the framework of Advanced Driving Assistance System (ADAS), the human driver’s intention needs to be recognised before the system intervenes with its assist functionality to facilitate drifting at the desired moment. We propose a method based on Bidirectional Long Short-Term Memory Network (Bi-LSTM) to interpret driver’s intention to start/exit drifting, utilising only basic driver inputs and vehicle state signals. Firstly, to comprehensively understand driver’s behaviour during drifting, we discuss the distinctive features in throttle and steering inputs and the corresponding vehicle acceleration signals during drift cornering and normal cornering, respectively. Next, two Bi-LSTM models are designed separately for the recognition of the ‘Intention to Start Drifting’ and the ‘Intention to Exit Drifting’. Then, these models are trained and evaluated through a data set that contains over 500 laps of driving collected from the racing simulator Assetto Corsa. To validate the proposed approach, test sets of different drivers, track layouts and car models are adopted. The proposed intention recognition models successfully reach an accuracy of over 90% in recognising the two concerned intentions and outperform other classification methods in comparison.