Feature-based Trajectory Anomaly Detection

The high availability of trajectory data in different fields makes it attractive to analyze and enhance its multiple practical applications. In particular, trajectory anomaly detection has a significant practical value, making it possible identifying trajectories that may indicate illegal and adverse activity in diverse areas such as surveillance, tracking devices, traffic, and people flow. This study presents a methodology to detect anomaly trajectories based on their morphology features. For that, we follow two stages: (1) comparative analysis of the performance of two descriptors to group similar trajectories, and (2) trajectory anomaly detection based on their similarities. We define Wavelet and Fourier transforms as trajectory descriptors to generate characteristics and subsequently detect anomalies ones. Our experiments emphasize the measure of the performance in the description of the coefficient feature space using unsupervised learning, specifically clustering techniques, to create subsets and identify irregular ones. The study's implications demonstrate that it is possible to use descriptors in trajectories for automatic anomaly detection and the use of unsupervised learning to segment required information. Our study's performance and comparative analysis have been demonstrated throughout multiple experiments. We present some quantitative results using synthetic data sets as well as qualitative analysis throughout two case studies considering real data sets that leave evidence of our contribution.