Outlier detection from vehicle trajectories to discover roaming events

Roaming, referring to the behavior of repeated observation of intended crime scenes before committing crimes, is a suspicious pattern often mentioned by experienced police investigators, but still remains a vague concept which needs to be well realized in video surveillance systems. This work first describes the scenario of roaming behaviors related to planned crimes, and then derives formal specifications for detecting suspicious roaming events from vehicle trajectories. Consequently, algorithms are designed for rapidly sorting out potential outliers and thoroughly examining their suspicious intention through circling activities, relative driving speed and time dispersion. Roaming trajectories and relevant trajectories are finally grouped into events and appropriately ranked. Furthermore, preliminary experiments and illustrative examples on synthetic data demonstrate the feasibility of our methods. The proposed approach enhances conventional vehicle plate analysis so that it becomes capable of discovering complex criminal behaviors and hence increasing investigation performance and decision quality.