A Multi-scale Fusion Method for Dense Crowd Counting

Abstract

AbstractWe propose a dense crowd detection network, called MSFNet, which can deal with highly dense crowd scenes, make accurate counting estimation and generate high-quality density maps by deep learning. The network is mainly composed of two main parts: the front-end network uses VGG-16 as the 2D feature extraction module, and the back-end network uses convolution networks with different sizes of convolution kernels instead of linking operations. The network is composed of convolution layers, which is an easy training model. We verify our network on two representative data sets (ShanghaiTech Data Set, UCF CC 50 Data Set), and the performance has been improved.In the first data set of ShanghaiTech, the root mean square error (MSE) decreased by 10%, and the mean absolute error (MAE) and root mean square error (MSE) of the second data set both decreased by about 6%.KeywordsMulti-scale fusionCrowd countingDense crowd