LAUSR

Driving behavior recognition is a challenging task that exploits the acceleration and angular velocity information of the vehicle collected by smartphone to identify various driving events. Traditional methods usually extract hand-crafted features from raw data, leading to under-explored temporal features of driving behaviors. To address the issue of hand-designed limitation for features, this paper proposes an end-to-end deep learning framework to automatically extract the features of driving behaviors. The mechanism behind our method is to model temporal features, capture salient structure features, and explore the correlation among the high-dimensional sensor data by fusing convolutional neural network (CNN) and recurrent neural network (RNN) with an attention unit. Moreover, a novel approach is introduced to build driving behavior dataset, which considers the effect of gravity in modeling smartphone sensor data. Subsequently, sensor data with device position independence is collected, and six types of driving events (straight driving, static, left turn, right turn, breaking, and acceleration) are annotated, which provides rich sensor information compared with other methods. The experimental results indicate that the proposed model outperforms other competing methods significantly, which possesses good generalization ability in the identification of driving behaviors.