Anomalous trajectory detection using support vector machines

One of the most promising approaches to event analysis in video sequences is based on the automatic modelling of common patterns of activity for later detection of anomalous events. This approach is especially useful in those applications that do not necessarily require the exact identification of the events, but need only the detection of anomalies that should be reported to a human operator (e.g. video surveillance or traffic monitoring applications). In this paper we propose a trajectory analysis method based on Support Vector Machines; the SVM model is trained on a given set of trajectories and can subsequently detect trajectories substantially differing from the training ones. Particular emphasis is placed on a novel method for estimating the parameter v, since it heavily influences the performances of the system but cannot be easily estimated a-priori. Experimental results are given both on synthetic and real-world data.