Counting Pedestrians Based on Weight-Minkowski-Dimension and Gaussian Process Regression

Abstract

A method is proposed to count the number of pedestrians based on Weight-Minkowski-Dimension and Gaussian process regression for fixed cameras surveillance. First of all, the crowd foreground was extracted using Gaussian mixture model, and then the Weight-Minkowski-Dimension, which count the boxes with weights that was calculated based on linear interpolation, was extracted in the binary image of foreground edge, and finally the number of crowd was predicted by and Gaussian process regression. And we evaluate the algorithm both in Fudan dataset and Pets2009 dataset. Experimental result shows that the Weight-Minkowski-Dimension not only responds the change of the crowd number, but also eliminates the influence of perspective distortions, thereby improves estimation accuracy. On the other hand, it performs better in crowded scene.