Article Preview
TopIn general, trajectory anomaly detection studies mainly include detection methods based on classification and clustering (Piciarelli et al. 2008, Yang et al. 2013, Li et al. 2007, Zhu et al. 2017, Kumar et al. 2017, Lv et al. 2017), methods based on distance and density (Lee et al. 2008, Liu et al. 2012b, 2013, San Román et al. 2019, Luan et al. 2017, Huang and Zhang 2019, Tang and Ngan 2016), and methods based on machine learning and pattern learning (Song et al. 2018, Ma et al. 2018, Bouritsas et al. 2019, Liu et al. 2020, Fu et al. 2020, Liatsikou et al. 2021).
The classification-based anomaly detection method first trains the classification model and then uses the pretrained model to judge whether the trajectory to be evaluated is anomalous. Piciarelli et al. (2008) suggested an anomalous trajectory detection method based on a support vector machine (SVM). In this method, every trajectory was described by a defined dimensional characteristic vector of the consistent samples of the original trajectory, and the trajectory classification was completed without prior information on the distribution of trajectory anomalies. However, classification-based methods need to annotate data for model training, which leads to a large amount of labor and time expenditure and reduces the usability of such methods to some extent. The clustering-based anomaly detection method clusters all trajectories into multiple groups. Dense classes are regarded as normal trajectories, and sparse classes are regarded as anomalous trajectories. Kumar et al. (2017) presented a two-phase clustering algorithm, in which trajectories were grouped by similarity measure while considering the trajectory direction information, and then the clustering results with fewer trajectories were judged as anomalous.