Dual Convolutional Neural Networks for Crowd Counting

Abstract

The core of crowd counting methods based on convolutional neural networks (CNN) is to construct a CNN model that can effectively mapping the crowd image to the density map, and then accurately estimate the number of crowd from the density map. The conventional CNN based methods only consider the mapping from crowd images to density maps without taking that from density maps to crowd maps and the impact of the mapping on model performance. A dual convolutional neural networks (DualCNN) is proposed to improve the accuracy of the mapping from crowd maps to density maps. DualCNN contains two mapping sub-models:1) a convolutional neural model that mapping a crowd image to a density map; 2) a dual convolutional neural model that mapping a density map to a crowd image. In the learning process, through the mutual influence of the two sub-models, the performance of the convolutional neural model on crowd counting is improved.Experimental results on the datasets of UCF_CC_50 and ShangHaitech show that the presented method can effectively improve the counting performance of CNN. Especially in the UCF_CC_50 data set, DualCNN reduces the MAE of MCNN and CSRNet by 15.6% and 15.8%, and MSE by 18.1% and 28.8%, respectively.