Detecting Anomalies from a Multitarget Tracking Output

Surveillance systems typically incorporate multitarget tracking algorithms for sequential estimation of kinematic states (e.g. positions, velocities) of moving objects in the surveillance domain of interest. This letter proposes an algorithm for online detection of anomalies in the motion and the count of objects, using the output of a multiobject tracking algorithm. The surveillance area is partitioned by a square grid and the kinematic states that fall inside each cell of the grid are modelled by a Poisson point process. During the unsupervised learning phase, the parameters of the Poisson point process are estimated for each cell. The testing phase is performed sequentially by threshold detection at a specified level of significance. The performance of the algorithm is illustrated using the Automatic Identification System (AIS) dataset in the context of maritime surveillance.