A knowledge-based approach for video event detection using spatio-temporal sliding windows

Scenario understanding from video stream plays an important role in many safety-critical application domains, especially if it is targeted at the autonomous aerial navigation and surveillance. Our contribution aims at enhancing the capability of unmanned aircraft systems to get a high-level description of the scene evolution from a video stream, by identifying events, thanks to the objects involved in these events. The work proposes a hybrid solution that merges data from the video tracking with additional semantic data: tracked objects are not only described by their typical information provided by tracking algorithms, but they are enhanced with semantic data, such as their geographical position, the interactions with other objects in the scene as well as their involvement in events occurring in a certain time interval. In particular, given a temporal window, a scene is depicted through the occurring events, the participating object tracks and the consequent evolution in terms of track movements.