Improving Deep Crowd Density Estimation via Pre-classification of Density

Abstract

Previous works about deep crowd density estimation usually chose one unified neural network to learn different densities. However, it is hard to train a compact neural network when the crowd density distribution is not uniform in the image. In order to get a compact network, a new method of pre-classification of density to improve the compactness of counting network is proposed in this paper. The method includes two networks: classification neural network and counting neural network. The classification neural network is used to classify crowd density into different classes and each class is fed to its corresponding counting neural networks for training and estimating. To evaluate our method effectively, the experiments are conducted on UCF_CC_50 dataset and Shanghaitech dataset. Comparing with other works, our method achieves a good performance.