Anomalous trajectory detection using a seq2seq auto-encoder

Abstract

Since the moving trajectory data has properties of a time series, we can effectively model the data employing a RNN-type model. The RNN model is one of the representative models in deep learning which has become a specialized model for time-series data by showing outstanding performance in various application fields that use the time-series data. However, unlike the common time-series data, the trajectory data is a spatio-temporal data that contains both spatial and temporal information. In other words, the trajectory data requires for the specialized methodology for feature extraction that utilizes both spatial and temporal dimensions. Therefore, simply constructing a RNN-type model is insufficient for feature extraction. Instead, we design an algorithm that efficiently captures distinctive high-quality features from the trajectories. For this purpose, we develop a Seq2Seq Auto-Encoder based model to extract complicated movement features and use extracted feature vectors to detect outlying trajectories.