基于全卷积神经网络的人群计数

Crowd counting based on fully convolutional neural network

基于从一张含有任意人群密度和任意视角的图像中准确地估计出其中的人群数目的目的,采用了全卷积神经网络先从图像中获得其人群密度图,然后对人群密度图上每个位置进行求和操作得到最终的人群数目的方法。所采用的全卷积神经网络不受输入图像的分辨率和视角的影响,同时,通过增加池化层层数,扩大网络的感受野,适应了图像中人头比较大的情况。所提出的算法在UCF_CC_50标准数据集上取得了最好的效果,进而验证了算法的高准确率和有效性。

In this paper,based on the purpose of accurately estimating the number of people from an image containing arbitrary crowd density and arbitrary perspective,the fully convolutional neural network(FCN)is used to obtain the crowd density map from the image,and then the values in the each position of the density map are summed to the final crowd count. The fully convolutional neural network is not affected by the resolution and perspective of the input image. In addition,by increasing the number of pooling layer,the receptive field of network is enlarged,which adapts to the large person head in the image. The proposed algorithm achieves the best performance on the UCF_CC_50 benchmark,which verifies its high accuracy and effectiveness.