A deep encoder-decoder network for anomaly detection in driving trajectory behavior under spatio-temporal context

Traditional trajectory anomaly detection aims to find abnormal trajectory points or sequences using data mining techniques. As a comparison, we focus on the evaluation of the anomalies of driving habits for different drivers based on their trajectory data. This is particularly important for the application of adjusting the amount of insurance in accordance with the driving behaviors. Instead of customizing rules for modeling various driving behaviors, we propose an end-to-end deep learning framework for driving trajectory anomaly detection, called STDTB-AD. Specifically, taking into account the fact that movement is spatial-temporal dependent, the study first partitions the whole road network into a series of spatial-temporal units, which have homogeneous properties of traffic flows. Then, the motion parameters (i.e., acceleration, speed, and direction) of driving trajectories falling within each spatio-temporal unit are calculated for representing context-aware features of drivers. This method is able to detect the deviation of movement from the normal traffic state on the spatial-temporal units. Finally, a variational autoencoder is utilized to quantify the abnormity degree of each driver according to the reconstruction probability of driving feature vector. Evaluations based on taxi trajectory data show that our model can consider both the spatial and temporal contexts for detecting driving behavior anomalies and achieve higher detection accuracy than traditional models based on either spatial or temporal context.