Taming Crowded Visual Scenes.

摘要

This report summarizes the progress made at Center for Research in Computer Vision at UCF in several distinct aspects of the project. We developed solutions to a number of important and challenging problems related to visual analysis of crowds and modeling of multi-agent interactions. The first category of solutions include behavioral analysis of crowds in videos captured through stationary as well as moving cameras. Second, we developed a novel technique to detect individuals in sparse crowds. This technique employs superpixels and iteratively improves the output of a generic underlying human detector. An advantage of the approach is that it outputs the exact segmentation of humans besides the more common bounding boxes. Third, we introduced and evaluated two new methods for analysis of extremely dense crowded scenes. The first approach deals with tracking of individuals in videos of high density crowd such as those depicting a marathon, while the other approach counts the number of individuals in an image of dense crowd. Finally, we developed a method that can detect anomalous behaviors in crowds using trajectories obtained through particle advection. Besides that, we also published the cover article in the prestigious Communication of ACM (CACM) and a book on modeling and analysis of crowds by Springer. All of research was conducted under the ARO's support during this project. The rest of the report describes these approaches in detail. There are five main sections: analysis of crowd behaviors, human detection in sparse crowds, visual analysis of extremely dense crowds, abnormal event detection, and finally we present details on the article and book published on visual analysis of crowds.