基于深度神经网络的人群计数方法  

Crowd counting method based on deep neural network

作　　者：郭利欣 李建军[1] Guo Lixin;Li Jianjun(School of Information Engineering,Inner Mongolia University of Science and Technology,Baotou,Inner Mongolia 014017,China)

机构地区：[1]内蒙古科技大学信息工程学院,内蒙古包头014017

出　　处：《计算机时代》2023年第8期5-10,15,共7页Computer Era

基　　金：国家自然科学基金项目(62066036);内蒙古自治区高等学校青年科技英才计划项目(NJYT22074);内蒙古自治区自然科学基金(2020MS06009)。

摘　　要：针对人群计数与人群密度估计任务中背景干扰、人群之间遮挡严重、上下文特征提取信息不足以及尺度变化巨大的问题,提出CA-CBAN模型由前端网络、CAN模块、CBAM模块和后端网络组成。将输入的图像传入前端网络进行初级特征提取;通过CAN模块进行多尺度特征融合;通过CBAM模块生成通道注意图与空间注意图;最后通过后端网络经过1×1的卷积获得密度图。该方法在ShanghaiTech数据集,UCF_CC_50数据集,JHU-CROWD++三个具有挑战的数据集上进行了实验验证,实验结果表明,所提方法的性能相较于CAN和其他的方法取得了较好的准确性和鲁棒性。Aiming at the problems in crowd counting and crowd density estimation tasks such as background interference,serious occlusion between crowds,insufficient information of context feature extraction and huge scale change,the CA-CBAN model composed of front-end network,CAN module,CBAM module and back-end network is proposed.The input image is transferred to the front-end network for primary feature extraction,multi scale feature fusion is performed through CAN module,channel attention map and space attention map are generated through CBAM module,and the density map is obtained through the back-end network with 1x1 convolution.This method is experimentally validated on ShanghaiTech dataset,UCF-CC_50 dataset and JHU-Crowd++dataset.The results show that compared with CAN and other methods,the proposed method achieves better accuracy and robustness.

关 键 词：人群计数 特征融合 人群密度估计 密度图 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]