TransDBC: Transformer for Multivariate Time-Series based Driver Behavior Classification
Jayant Vyas, Nishit Bhardwaj, Bhumika Bhumika, Debasis Das
Abstract
Improving driving safety by monitoring driver behavior is an excellent example of Advanced Driver Assistance Systems (ADAS). This paper proposes an end-to-end transformer-based driver behavior classification framework named Trans-DBC. It calculates driver behavior from the multivariate time-series smartphone telematics data by learning short and long-range temporal dependencies effectively and accurately, unlike prior data-driven deep learning models that capture information locally via convolutional or recurrent structure in an iterative manner. The extensive human-in-the-loop study on a publicly available UAH-DriveSet dataset shows that the proposed technique can classify unsafe driving behavior with a 96% of average weighted precision, recall, F-measure, and 95.38% accuracy. Our suggested model outperforms baselines and state-of-the-art models on the UAH-DriveSet dataset and five multivariate time-series datasets in driving behavior analysis.