Abstract: The purpose of this paper is to accurately estimate the crowd density in real scenes based on image information from arbitrary perspective and arbitrary crowd density.However,crowd counting on static images is a challenging problem.Due to the perspective distortion and the crowd crushes caused by the projection from 3D space into 2D space,it is difficult to distinguish the difference between individual and individual and the difference between individual and background.To this end,this paper proposed a flexible and efficient scalable modularized convolutional neural network (CNN) architecture.The network allows to directly input images with arbitrary size and resolution and it does not require additional computational changes in view information.Each module of the architecture employs a multiple column structure with different convolution kernels,which can be used to fit individual information of different distances.The proposed module also combines the feature information of the front and rear two layers,reducing the decrease loss of the accuracy caused by the vanishing of the gradient.Experiments show thatthe accuracy of proposed method is increased by 14.58% and 40.53%,and the root mean square error is reduced by 23.89% and 33.90% respectively on ShanghaiTech PartA and PartB datasets compared with the state-of-the-art MCNN methods.

Key words: Convolutional neural network, Crowd counting, Density maps, Feature fusion, Scalable module