Driver Behavior Classification: A Novel Approach Using Auto-Encoders and Motif Extraction

Abstract

Driver’s behavior is expressed by the intentional or unintentional actions the driver performs while driving a motor vehicle. This behavior could be influenced by several factors such as driver’s fatigue, drowsiness, vehicle surroundings, or distraction state. Monitoring, analyzing and improving driver’s behavior can reduce traffic collisions and enhance road safety. Several approaches have been followed for the detection and identification of driver’s behavior. Conventional time-series analysis applies forecasting analysis methods for driver’s behavior detection, assuming that data are stationary and ergodic; otherwise data preprocessing is mandatory. Rule-based and deep learning approaches have succeeded to mine dynamical characteristics of driving time series data. However, they have some challenges, including the selection of efficient architectures and corresponding hyper-parameters, as well as slow training and limited labeled data. In this study, we propose a motif-based approach for categorizing driver behavior as normal or abnormal, using the UAH-DriveSet dataset. Our methodology entails the selection of relevant features, which are encoded using an auto-encoder model, followed by the conversion of the encoded data into an alphabet representation through quantization. Unique patterns of varying lengths are then extracted, and an AdaBoost classifier is utilized for behavior classification. Extracted motifs capture significant patterns, which enables to achieve higher accuracy in classification. The obtained results demonstrate the effectiveness of the proposed approach in accurately categorizing driver behavior, which can significantly contribute to the advancement of intelligent transportation systems and the enhancement of road safety.