UT-ATD: Universal Transformer for Anomalous Trajectory Detection by Embedding Trajectory Information

Abstract

Due to the development of the transportation industry, a large amount of trajectory data is pouring into the Internet all the time. Based on these trajectory data, anomalous trajectory detection technology provides great support for traffic safety assurance and traffic risk prediction. Most existing anomalous trajectory detection methods are based on trajectory's physical characteristics or representation learning, and they achieve good performance in a few scenarios. But they still face the following problems. (1) The imperfect utilization of trajectory points. (2) The sparsity of trajectory data, which leads to generalization issues. (3) Longer model training time consumed, which can't adapt to the large amount of trajectory data generated every day. To solve the above problems, we propose a novel anomalous trajectory detection model based on Universal Transformer, called UT-ATD. UT-ATD captures the information of trajectory positions by learning trajectory embedding for classification. UT-ATD has a faster training speed, relatively few model parameters, and sufficient portability, which are ideal for the realistic scene requirements. Our model achieves state-of-the-art performance in most aspects, and its effectiveness is verified by a series of experiments on the real-world taxi trajectory dataset.