A CNN-RNN Neural Network Join Long Short-Term Memory For Crowd Counting and Density Estimation

Abstract

Crowd counting and density estimation is a challenging task in the field of computer vision. Most of existing methods of this task are based on convolutional neural network (CNN), which have achieved good results in low-density scene. Usually, people who are far away from the camera appear to be denser and smaller, while those who are close to the camera are more sparse and larger, therefor, structure contains only CNN gives the poor performance in some high-density crowd scene because of the uneven distribution of the crowd through camera. To address this problem, this paper designs a CNN-RNN Crowd Counting Neural Network (CRCCNN), which introduces Long Short-Term Memory (LSTM) structure, we use CNN structure to extract the features of the whole image, and use the LSTM structure to extract the contextual information of crowd region. Since LSTM has a good memory of the input information of sequential samples, it can predict the crowd density very well even for the high density population. We perform our experiments on different datasets and compare with other existing methods, which achieve the outstanding results and demonstrate the effectiveness performance of CRCCNN.