基于卷积神经网络的人群计数算法研究  

Research on Crowd Counting Algorithm Based on Convolution Neural Network

作　　者：向飞宇 张秀伟[1] XIANG Fei-yu;ZHANG Xiu-wei(School of Computer Science,Northwestern Polytechnical University,Xi’an 710129,China)

机构地区：[1]西北工业大学计算机学院,陕西西安710129

出　　处：《计算机技术与发展》2021年第7期42-46,共5页Computer Technology and Development

基　　金：国家级大学生创新训练项目(201910699020)。

摘　　要：随着城市监控网络的完善,对人群图像的计数处理正产生巨大价值。传统人群计数方法存在准确度低,无法处理高遮挡图像,受光影影响大等问题。卷积神经网络在人群计数上表现出色,但仍存在精确度较低,无法排除背景图像干扰等问题。为提高对复杂人群图像的感知能力,减少背景区域对统计的影响,并同时生成人群密度特征图像,在卷积神经网络的基础上增加空间与通道注意力模型,对不同通道和不同位置的图像赋予不同的权重以增加目标区域的影响力,同时更换全连接层为上采样层,输出与输入图像大小相同的人群密度特征图像。实验中使用ShanghaiTech数据集以及NWPU-Crowd数据集进行训练与测试,在与MCNN、CSRNet等网络的比较结果中显示,使用了注意力模型与全卷积神经网络的算法在平均绝对值误差与均方误差两项数据上有较好的结果,表示该算法在高密度高遮挡的人群图像计数上有着更高的精确度。As the growing of urban monitoring network,the counting of crowd image is proved to be of great value.Traditional methods of crowd counting have problems like low accuracy,inability to deal with high occlusion images and being greatly affected by light and shadow.Convolution neural network performs well in crowd counting,but problems like rather low accuracy and being influenced by background area still exist.To improve the perception of complex crowd image,decrease the influence of background area and create crowd density image in the same time,spatial-wise and channel-wise attention model are added to convolution neural network in order to give more weight to target area.Full connect layers are replaced by up-sampling layers to output crowd density image by the same size of the original input image.ShanghaiTech dataset and NWPU-Crowd dataset are used to train and validate the network,and the comparison among networks like MCNN and CSRNet shows that the proposed algorithm with attention model and fully convolutional network has better results in the mean absolute error and mean square error data,indicating that it has a higher accuracy in the high-density and high-occlusion crowd image counting.

关 键 词：人群计数 全卷积神经网络 注意力模型 扩张卷积 特征提取 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]