Driver Profiling Using Long Short Term Memory (LSTM) and Convolutional Neural Network (CNN) Methods

Abstract

Driver profiling has a major impact on traffic safety, fuel consumption and gas emission. LSTM and CNN based neural network models were developed to classify and assess bus driver behavior characterized by deceleration, engine speed pedaling, corner turn and lane change attempts. Deceleration, engine speed and corner turn test scenarios were performed on concrete paved test track while lane changing tests were conducted on a commercial asphalt highway. Despite the majority of studies relying on image, vehicle data and additional sensor fusion, here only the data streams received from vehicle CAN Bus system were used to train the proposed network architectures. After parsing the data into meaningful characteristic parameters, different LSTM and CNN architectures were trained by varying the number of layers, neurons and epoch number. Both LSTM and 1D-CNN networks resulted in comparable success rates. CNN architecture indicates better performance indices for identification of aggressive driving compared to LSTM network for behavioral modelling.