Crowd counting and density estimation via two-column convolutional neural network

Abstract

This paper proposes a Two-Column Convolutional Neural Network (TCCNN) to estimate the density and count of both sparse and highly dense crowd. The architecture of TCCNN derives from VGG-16 and Alexnet. We concatenate parts of these two networks to output the estimated density map and Gaussian Kernel is employed to generate the true density map as ground truth for training. Through integral on the entire density map, the number of people within the image can be obtained. We test the proposed method on such challenging datasets as UCF_CC_50, Shanghaitech and UCSD, to which different data augmenting methods are applied. The results show that our method is of competitive performance in comparison with other state of the art approaches.