Predictive Visual Saliency Model for Crowd Videos

Abstract

Understanding saliency in a scene is vital as it helps to understand our human visual attention. There is a constant need for new, innovative datasets that cover a particular category of scenes from all angles. Crowded scenes are an example of such specialized scenes. This thesis introduces a new and dynamic crowd video dataset while detailing the eyetracking experiment through which it was procured. The dataset is annotated into three categories of crowd density and evaluated on existing general saliency model to get a benchmark on their performance and suggested improvements for the future design of crowd saliency model.