Abstract: In an attempt to improve the counting accuracy of crowd images, a multi-task multi-level convolutional neural network is designed in this paper. Multi-level neural network is constructed with combinations of convolution and upsampling. The network has the advantages of combining the detailed information extracted by the shallow layer and the high-level semantic information extracted by the deep layer. On this basis, multi-task learning is used to improve network performance. Multi-task learning is divided into two parts, namely, crowd density estimation and crowd density classification. The high-resolution layer of the network is connected to crowd density estimation task, and the deep layer is connected to crowd density level classification task. A new loss function is defined by merging the losses of two tasks. The experiment is carried out on the crowd public data set ShanghaiTech, WorldExpo’10 and UCF_CC_50. The experimental results reveal that the proposed network has better accuracy and robustness in crowd counting.

Key words: crowd counting, Convolution Neural Network（CNN）, multi-task learning, crowd density estimation, crowd density classification

摘要： 为了提升人群图像的计数精度，设计一种多层级多任务深度卷积网络。多层级神经网络由卷积和上采样的组合方式构成，该网络的优点在于结合浅层网络提取的细节信息和深层网络提取的高阶语义信息。在此基础上，使用多任务学习的方法提升网络性能，多任务学习分为两个部分：人群密度估计任务和人群密度等级分类任务。网络的高分辨率层与人群密度估计任务相连，网络的深层与人群密度等级分类任务相连。将两个任务的损失融合并构成新的损失函数。实验在人群计数公共数据集ShanghaiTech、WorldExpo’10和UCF_CC_50上进行，实验结果表明，该网络在人群计数上具有较好的准确率和鲁棒性。

关键词: 人群计数, 卷积神经网络, 多任务学习, 人群密度估计, 人群密度等级分类