The widespread use of video surveillance cameras makes impossible the human analysis of all the video data. The current study searches for a solution to this problem by detecting unusual events (in real time) and/or providing a query mechanism, in order to find a specific situation.

In order to detect unusual situations, there are two approaches. Either there is a known model for unusual situations, in which case it is enough to build a « video parser », or the system learns the usual situations during an observation period, after which anything which is not usual is labeled as unusual.

The system developed in this study uses the second approach. Our main interest concerns the trajectories and the various distances and classification methods for trajectories. Moreover, the user should have the choice for the discriminant factors between trajectories (speed, orientation, etc.)
Succinct information about trajectories, as well as some characteristics of the objects (e.g., size, color, etc.) are saved and indexed, so that they may be easily retrieved on demand.