The research in era of driving behavior is promising and of great importance for traffic safety. The lane changing behaviors are primary driving tasks on traffic While relatively less attention has been paid into the anomalous lane changing behaviors and deep understanding of them was lacking. The Attention based Gate Recurrent unit Autoencoder (ATGRU-AE) and Isolation Forest were used as a combination to identify the anomalous lane changing events from the Shanghai naturalistic driving study datasets. Isolation Forest is employed to obtain the anomaly scores based on the latent and compact nonlinear representations of the time-series data extracted by the GRU-AE model trained. By means of kernel density estimation and temporal attention mechanism, we got deep understanding of the characteristics of the anomalous lane changing behaviors in a personalized way. The results show the anomalous rate for a special driver is about 2%~8%. Anomalous lane-changing samples show display distinct different over variables probability distribution compared to the whole samples group and normal groups. Besides the attention weights of time-series data over different time steps are uneven and the intermediate ones receive less weights. Our study contributes to getting better in sight of the spatio-temporal characteristics of anomalous lane-changing behaviors.

CICTP