One-Dimensional Convolutional Neural Network Model for Abnormal Driving Behaviors Detection Using Smartphone Sensors

Abnormal driving behavior detection (ADBD) is essential to improving driving safety. Traditional approaches usually extract features manually, resulting in insufficient exploration of deep features of driving behavior. To address the limitation of shallow feature-based approaches, we propose a one-dimensional convolutional neural network (1D CNN) model for ADBD and transforms it into the time series multi-classification. Firstly, we construct a dataset of driving behavior using smartphone sensors, and five fine-grained abnormal driving behaviors (Hard braking, Weaving, Swerving, Quick turn, Quick U-turn) are defined and labeled by analyzing the abnormal driving feature patterns. Then, we need to map the smartphone coordinate to the vehicle coordinate, and apply the low-pass filtering on the input data for high-frequency noise reduction. Finally, we train the dataset with the 1D CNN model for feature extraction and classification. The experimental results show that the proposed 1D CNN model efficiently achieves multi-classification of abnormal driving behaviors with an average accuracy of 97%, significantly better than the traditional algorithm of k-nearest neighbor and support vector machine.