编码-解码多尺度卷积神经网络人群计数方法  

Encoding-Decoding Multi-Scale Convolutional Neural Network for Crowd Counting

作　　者：孟月波[1] 纪拓 刘光辉[1] 徐胜军[1] 李彤月 MENG Yuebo;JI Tuo;LIU Guanghui;XU Shengjun;LI Tongyue(School of Information and Control Engineering, Xi’an University of Architecture and Technology, Xi’an 710055, China)

机构地区：[1]西安建筑科技大学信息与控制工程学院,西安710055

出　　处：《西安交通大学学报》2020年第5期149-157,共9页Journal of Xi'an Jiaotong University

基　　金：国家自然科学基金资助项目(51678470);陕西省教育厅专项科研计划资助项目(18JK0477);西安建筑科技大学基础研究基金资助项目(JC1703)。

摘　　要：针对基于多列卷积神经网络的人群计数方法存在的多尺度特征信息丢失、融合不佳以及密度图质量不高等问题,提出了一种编码-解码结构的多尺度卷积神经网络人群计数方法。编码器采用多列卷积捕获多尺度特征,通过空洞空间金字塔池化扩大感受野并减少参数量,保留尺度特征和图像的上下文信息;解码器对编码器输出进行上采样,实现高层语义信息和编码器前端低层特征信息有效融合,从而提升了密度图的输出质量。为增强网络对计数的敏感性,在以往像素空间损失的基础上考虑了计数误差,提出了一种新型损失函数。采用Shanghai Tech、Mall以及自建数据集进行了对比实验,结果表明:与之前最优方法相比,所提方法在Shanghai Tech数据集Part_A部分的平均绝对误差和均方误差分别降低了8.3%和21.3%,Part_B部分分别降低了12.9%和12.0%,Mall数据集分别降低了15.1%和23.8%,自建数据集分别降低了13.5%和7.1%;在不同人群场景下,所提方法的人群计数准确性和鲁棒性均优于其他对比方法的。Aiming at the problems of multi-scale feature information loss,poor fusion and low quality of density map in the crowd counting method based on multi-column convolutional neural network,a new crowd counting method is proposed based on encoding-decoding multi-scale convolutional neural network.The encoder part adopts multi-column convolution to capture multi-scale features,expands the receptive field and reduces the amount of calculation via atrous space pyramid pooling,and retains the multi-scale feature and the context information of the image.The decoder part upsamples the encoder output to achieve effective fusion of the features with rich high-level semantic information and the features with rich low-level detail information to improve the output quality of the density map.To enhance the sensitivity of the network to counting,a new loss function is proposed by considering the previous pixel space loss and the counting error.Contrast experiments with previous methods on Shanghai Tech,Mall,and self-built datasets are conducted,and it is found that the mean absolute error and mean square error of this method on part_A of Shanghai Tech dataset are 8.3%and 21.3%lower than the previous optimal method,and 12.9%and 12.0%lower in part_B of Shanghai Tech dataset.The mean absolute error and mean square error decrease by 15.1%and 23.8%for the Mall dataset,and decrease by 13.5%and 7.1%for the self-built dataset.The experimental results on Shanghai Tech,Mall and self-built datasets show the higher accuracy and better robustness of the proposed method than the traditional methods.

关 键 词：人群计数 编码-解码结构 多尺度 空洞空间金字塔池化 计数误差 损失函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]