On Temporal Feature Engineering for Driver Behaviour Prediction

Driver behaviour modelling is a pivotal field of research that seeks to understand the complex and ever-changing driving behaviours on the road, leading to improvement of road safety, reduction in air pollution, and more effective vehicle performance. Thanks to advancements in sensor technology and machine learning (ML) algorithms, capturing and evaluating driver behaviour patterns become significantly more accessible. However, the accuracy of ML models relies heavily on data quality, so the role of feature extraction techniques in delivering high-quality inputs is crucial. The primary contribution of this research is to evaluate and compare various feature extraction techniques with the aim of enhancing data quality and ultimately improving the prediction of driver activities. We also designed, implemented, and evaluated two novel feature extraction approaches called Statistical Lag Feature (SLF) and Behaviour Feature (BF) methods to extract features from raw sensor data. The experimental findings unequivocally demonstrate that the SLF and BF techniques exhibit significantly improved prediction accuracy of ML models compared to Fast-Fourier Transform (FFT), Convolutional Neural Networks (CNN), Statistical, and Cross-Correlation methods.