Pedestrianly event detection using grid-based features

Video surveillance systems in public areas are grown rapidly for safety and security; therefore, the number of monitors becomes too large to watch by human. Automatic event detection system becomes more important. A trouble of surveillance camera in pedestrianly areas is that position of camera is too far or too close to the target objects and it compromises detection performance. In order to limit effects of camera positions, this paper proposes an event detection framework using grid-based features, which is a combination of localized information and event rules. Relationship between grid resolution and accuracy performance of event detection is studied. Grid-based features are tested on Neural Network and SVM classifiers. Experimental results show that grid-based features perform better than non-grid features. Performance of learning machines is also related to event types and grid size. The larger grid size is appropriate for the farther camera position.