Uma abordagem baseada em regras para detecção automática de eventos pré-definidos em vídeos.

Abstract

The main objective of this work is to develop a system for automatic detection of predefined events in videos. The automatic monitoring of videos is a relevant research topic when considering the large number of applications where it is impossible to have continuous human monitoring. As examples, it can be cited road traffic situations, scenarios involving access control to people or vehicles, arrival area at airports, among others. The proposed system is composed of an architecture designed to detect events that involve interactions between moving objects and between these objects and contextual elements in the scene, represented as 2D shapes. A distinctive feature of the system is the modeling of events and scenarios through a logic specification using rules and facts in CLIPS - an environment for building expert systems. Low-level modules of the architecture uses computer vision algorithms such as video background subtraction and tracking to update database of facts of interest, while the inference engine of CLIPS uses this information to detect predefined high-level events. The system is illustrated with cases of crossroad traffic situations involving violations of traffic rules between cars, pedestrians, including violations of signal lights of semaphores. Tests were conducted with videos processed off-line and the results of the event detection system are compared with ground truth manually annotated, producing promising results.