A Feature-Fusion-based Multi-column Convolutional Neural Network for Crowd Counting and Density Estimation

Due to the scale change caused by perspective distortion, the automatic estimation of crowd density in images with high density population is still a extremely difficult task. To address this problem, Feature-Fusion-based Multi-column Convolutional Neural Network (F2MCNN) is proposed to perform accurate crowd count estimation and provide high-quality density maps. Multi-column convolutional neural network are used to extracted multi-scale features distributed in different regions in a single crowd image. In this paper, the intermediate network nodes of different columns are connected by hopping to fuse multi-scale features and improve the perception ability of images at different scales. F2MCNN is evaluated on four datasets and it achieves better mean absolute error and mean squared error performances compared with MCNN.