Distributed Data Analysis for Driving Behavior Estimation Using Auto ML and Deep Learning

The identification of driving behavior is a crucial factor in road accident prevention. Two drivers’ control features (i.e., 3-axis accelerometer and gyroscope signals) are used for the classification process to identify the driving behaviors. The data in the public dataset (Mendeley data) are labeled as ‘normal driving behavior’ and ‘aggressive driving behavior’ based on the so-called ‘jerk profile’ and ‘yaw rate’. The jerk profile describes the rate of acceleration changing over time, while the yaw rate is the rotation of a vehicle around its vertical axis. A preprocessing step deals with Allan variance analysis performed on the accelerometer and gyroscope collected data to evaluate the noise characteristics that could affect the data. Different classification models are tested to identify a person's driving style accurately. The first approach belongs to AutoML, which is based on PyCaret, and the second approach deals with Deep Neural Network DNN, which utilizes AutoKeras. The noise characteristics of the data indicate that the dominant noise for short window times is white noise and that it satisfies acceptable standards. The algorithms successfully identify the drivers’ behavior with accuracies of 0.8341 for the Extra Trees Classifier and 0.8305 for deep learning architecture. The proposed approach generates good results and demonstrates the feasibility of developing a system based on Artificial Intelligence to predict driving behavior when the driving data is automatically collected.