Estimation of crowd density based on deep convolutional neural networks

摘要

Crowd density estimation is a significant research contents in the intelligent surveillance system, which is a valid method for public security and flows management. But the existing methods are not able to accommodate the demand of the practical applications, in virtue of crowd occlusions, perspective distortions and variable weather. In addition, most existing methods always use general the hand-designed features, which don't have enough representation ability of the crowd. To address these problems, in this paper, we propose a deep convolution neural networks (DCNN)-based method to estimate the crowd density in natural scenes. Firstly, we divide the crowed image into several image patches according to the criterion of the mean height of the adult pedestrian, which overcome the impact of perspective distortion on the pedestrian images. Secondly, the deep convolutional neural network has been designed. The DCNN is used to extract crowd features by different convolution kernels on the pedestrian image. The learned crowd features are employed to estimate crowd density. We test our approach on three different data sets, the experimental results show our DCNN have the accuracy and robustness in the different scenes.