Crowd counting by feature-level fusion of appearance and fluid force

Abstract

Crowd counting is a research hotspot for video surveillance due to its great significance to public safety. The accuracy of crowd counting depends on whether the extracted features can effectively map the number of pedestrians. This paper focuses on this problem by proposing a crowd counting method based on the expression of image appearance and fluid forces. Firstly, Horn-Schunck optical flow method is used to extract the motion crowd. Secondly, based on the motion information of crowd, pedestrians in different directions are distinguished by the k-means clustering algorithm. Then, image appearance features and fluid features are extracted to describe different motion crowd. The image appearance features are gained by calculating the foreground area, foreground perimeter and edge length. The gravity, inertia force, pressure and viscous force are taken as the fluid features. Finally, two kinds of features are combined as the final descriptor and then least squares regression is used to fit features and the number of pedestrians. The experimental results demonstrate that the proposed crowd counting method acquires satisfied performance and outperforms other methods in terms of the mean absolute error and mean square error.