Research and Realization of Crowd Density Estimation Based on Video

Abstract

Automated estimation of crowd density and early warning for overcrowded situation are essential and valid approaches for public security management. This paper builds a system to detect and monitor the level of congestion aimed at crowded scenes. Obtaining the crowd foreground by Gaussian Mixture Background Modeling to eliminate the bad effect the changing of complex background does to the detecting result. Extracting the textural features of the crowd foreground to avoid the misjudgment issue caused by people covering each other while purely using pixels statistics method to analyze crowd foreground. This paper uses GRNN neural network to train test samples to get the density level classifier. The accuracy of high density for the experimental scene in this paper is up to 92%. Proven the performance is relative good and the system this paper built is practicable and valid.