Abstract
It is evident that crowd counting is one of bottlenecks for crowd-related computer vision theory and applications such as surveillance. Since accuracy of estimating crowd size dominantly depends on the performance of motion detection of pedestrians, this paper attacks the challenging problem mainly by proposing a motion segmentation method based on flow field texture representation. Firstly, the motion crowd and background are represented as different texture images by employing line integral convolution. Then information entropy is introduced to quantify the textures as different values so that the texture images can be segmented; further an optimal threshold is obtained via Otsu method to segment the binarization entropy image. Finally, the area of motion foreground pixels is calculated for each image in a crowd motion video. The size of the crowd is estimated by least squares fitting using abundant datum of foreground pixels’ area and the number of individuals in a crowd. Experimental results demonstrate that the proposed crowd counting method outperforms background subtraction, Gaussian mixture model and optical flow-based methods in terms of mean absolute error and mean relative error.
Original language | English |
---|---|
Pages (from-to) | 871–883 |
Number of pages | 13 |
Journal | Machine Vision and Applications |
Volume | 26 |
Issue number | 7-8 |
Early online date | 14 Aug 2015 |
DOIs | |
Publication status | Published - Nov 2015 |
Keywords
- image understanding
- crowd counting
- flow field visulization
- line integral convolution
- information entropy
- Otsu segmentation
- least squares fitting