Abstract: Distracted driving such as phone use during driving is risky, as it increases the probability of severe crashes. Detecting distraction using Naturalistic Driving Studies was attempted in existing studies, and most of them used facial motions, which would be highly influenced by light conditions and algorithm effectiveness, still could not fully indicate auditory and physical distractions. This study aims to optimize Long Short-Term Memory (LSTM) model for phone usage detection based on vehicle dynamics sensor data from Shanghai Naturalistic Driving Study (SHNDS), China. A total of 1244 phone use events were extracted from videos of SH-NDS, and analyzed against focus driving baseline. Performance attributes included speed, longitudinal acceleration, lateral acceleration, lane offset, and steering wheel rate. Their mean, standard deviation, and predicted error (PE) were calculated, and derived 15 indicators. A Bidirectional layer and attention mechanism were added to the LSTM model for higher accuracy. Results showed that besides the mean and standard deviation of steering wheel rate, all the other 13 indicators were significant and effective in the model. The Bidirectional Long Short-Term Memory (BiLSTM) model reached a promising result of approximately 91.2% accuracy using 5-fold cross validation, which was better than other machine learning methods such as recurrent neural network, support vector machine, k-nearest neighbor, and adaptive boosting. This Bi-LSTM model with attention mechanism could potentially be applied in advanced driving assistant systems to warn driver and reduce phone involved distracted driving.
Xuesong Wang, Rongjiao Xu, Siyang Zhang, Yifan Zhuang, Yinhai Wang. Driver distraction detection based on vehicle dynamic using naturalistic driving data. Transportation Research Part C. 2022.