A Convolutional Neural Network with Background Exclusion for Crowd Counting in Non-uniform Population Distribution Scenes

The crowd counting in public places is a wildly concerned issue in the fields of public safety, activity planning, and space design. The current crowd counting methods are mainly aimed at the situation that the crowd is full of the whole scene, which cannot be applied to practical applications due to the actual crowd is non-uniform distributed in the scene. The complex background caused by non-uniform population distribution affects the accuracy of crowd counting. Therefore, we propose a convolutional neural network with background exclusion for crowd counting. Firstly, we divide the image into blocks and then use the residual network to determine whether each block contains crowd, to eliminate the clutter background area and avoid the background interference to crowd counting. Secondly, we use the dilated convolution and asymmetric convolution to estimate the crowd density map of image blocks containing crowd. Finally, the crowd density map of all crowd areas is integrated to obtain the crowd counting results of the whole scene. We collect some images of more general scenes, such as the crowd is only a part of the whole image, and construct Non-uniformly Distributed Crowd (NDC 2020) dataset. We conduct experiments on ShanghaiTech datasets and NDC 2020 dataset. Experiment results show that our method is superior to the existing crowd counting methods in the scene of non-uniform population distribution.