Dynamics analysis for local crowd state without foreground segmentation

Considering that a static background model can not be used to precisely confirm the crowd state in a complex scene of surveillance video due to its poor adaptability, a scheme for dynamics analysis of local crowd state without foreground segmentation is proposed. The scheme handles local blocks in consecutive frames in the space-time domain as a linear dynamic system (LDS), and employs the mixture dynamic texture algorithm to classify them to estimate the crowd density; uses a main path tracking method to evaluate the crowd velocity; models the LDS by partial differential equations to describe the variation relation between the density field, velocity field and flow quantity field. The experimental results show that the proposed scheme can be used to perform quantitative analysis on the crowd state, as well as on the changing trend. The result of state analysis can be used to detect the anomaly events in a crowd exactly.