An Eigenvalue Approach to Detect Flows and Events in Crowd Videos

摘要: Analysis of flows in crowd videos is a remarkable topic with practical implementations many different areas. In this paper, we present wide overview along our own approach to problem. Our treats the difficulty flow analysis by distinguishing single versus multiple scene. Spatiotemporal features two consecutive frames are extracted optical create three-dimensional tensor, which retains appearance and velocity information. Tensor’s upper left minor matrix captures intensity structure. A normalized continuous rank-increase measure for each frame calculated generalized interlacing property eigenvalues these matrices. essence, values put through knowledge existing flows. Yet they do not go into effect desirably due estimation error some other factors. proper set degree polynomial fitting functions decodes their existence. But how can estimate that set? Its detailed study performed. Zero flow, flows, interesting events detected as basis using thresholds on values. Plausible mean outputs recall rate (88.9%), precision (86.7%), area under receiver operating characteristic curve (98.9%), accuracy (92.9%) reported from conducted experiments PETS2009 UMN benchmark datasets make clear visible method gains high-quality results detect terms both robustness potency.