A Customized Kernel-based Deep Stacked Autoencoder with Optimal Ensemble Recurrent Neural Network Architectures for Driver Status Prediction

Road safety is crucial to everybody who is a road user, pedestrian, and/or driver. Unfortunately, according to the World Health Organization (WHO), 1.19 million deaths and 50 million injuries are caused by road traffic accidents each year. The significant figures in traffic deaths and injuries are unacceptable and preventable. Modern driver monitoring systems have become more applicable and reliable in real-world scenarios, using cutting-edge technologies, the Internet of Things, and machine learning algorithms. Leveraging automated monitoring and predictive analytics for driver status is a promising preventive strategy. This study highlights the critical issues of drowsy, distracted, and stressed driving, identified as prevalent undesired driving behaviors (UDBs). The objective of this research is to bridge gaps in capturing distinctive features from individual training datasets, which include both behavioral and environmental variables relevant to each driver, and integrate these into the machine learning framework. Moreover, there has been limited focus on the necessity for timely predictions of UDBs, considering their varying onset times. To address these challenges, we propose a customized Kernel-based Deep Stacked Autoencoder (KDSAE) that aims to extract both common and unique characteristics from an individual’s driving patterns alongside environmental influences, ultimately enhancing the prediction accuracy for driving behavior. An ensemble learning strategy is used to design optimal prolonged driver status predictions for three driving behaviours: drowsy, distracted, and stressed driving. The strategy involves fusing three recurrent neural network architectures: a long short-term memory network, a gated recurrent unit, and a traditional recurrent neural network. Three benchmark datasets (related to stressed driving, drowsy driving, and distracted driving) are selected for evaluating algorithms, and ablation experiments revealed the effectiveness of individual components. The proposed prolonged prediction models for stressed, drowsy, and distracted driving predictions achieved average root-mean-square error (RMSE) of 0.176, 0.173, and 0.207, respectively.