3D virtual modeling for crowd analysis

This paper presents a method to analyze crowd with computer vision techniques in virtual environments. To overcome the difficulty of obtaining video evidence in hazard situations, or, to meet the demand of big data for machine learning methods, we attempt to use virtual models to simulate actual ones. To prove the reliability of virtual crowd models we simulated in three situations where people walk normally, somebody runs fast, and crowd gathered, we collected corresponding real-life videos. Then, we adopted techniques of pedestrians detection and optical flow extraction to examine both virtual and actual crowd models. Finally, we compared the results of model analysis and found that those two kinds of models both meet the requirements. At this stage, we proved that the method we proposed could find its use in video surveillance, 3D reconstruction, and any other demands for the combination of computer vision and computer graphics.