密集场景下基于多尺度特征聚合的人群计数方法  

Crowd Counting Based on Multi-scale Feature Aggregation in Dense Scenes

作　　者：刘培刚[1] 孙洁[1] 杨超智 李宗民[1,2] LIU Peigang;SUN Jie;YANG Chaozhi;LI Zongmin(School of Computer Science and Technology in China University of Petroleum(East China),Qingdao,Shandong 266580,China;Shengli College of China University of Petroleum,Dongying,Shandong 257061,China)

机构地区：[1]中国石油大学(华东)计算机科学与技术学院,山东青岛266580 [2]中国石油大学胜利学院,山东东营257061

出　　处：《计算机科学》2023年第9期235-241,共7页Computer Science

基　　金：国家重点研发计划(2019YFF0301800);国家自然科学基金(61379106);山东省自然科学基金(ZR2013FM036,ZR2015FM011)。

摘　　要：密集场景下个体尺度存在巨大差异,目标个体尺度不一导致人群计数精度不高。针对这一问题,提出了一种密集场景下基于多尺度特征聚合的人群计数方法。该方法研究不同特征层级对不同尺度个体的特征信息表示能力,通过层级连接充分获取多尺度特征;同时,提出了一个多尺度特征聚合模块,采用多列具有不同扩张率的空洞卷积,通过动态特征选择机制自动调整感受野,以有效提取不同尺度个体的特征。该方法能够在保留小尺度个体特征信息的基础上进一步扩大感受野,增强大尺度个体的检测能力,使其更好地适应人群个体的多尺度变化。在3个公共人群计数数据集上进行了实验,实验结果表明,所提模型在计数准确性上有了进一步的提高,其中在ShanghaiTech数据集Part_A上MAE为51.21,MSE为83.70。Individual scales vary greatly in dense scenes,and the varying scales of target individuals lead to poor crowd counting accuracy.To address this problem,the crowd counting method based on multi-scale feature fusion in dense scenes is proposed.The method investigates the ability of different feature layers to represent feature information for individuals at different scales,with adequate access to multi-scale features through layer connections.At the same time,a multi-scale feature aggregation module is proposed,which uses multiple columns of dilated convolution with different expansion rates,and automatically adjusts the perceptual field through a dynamic feature selection mechanism to effectively extract features of individuals at different scales.The method can further expand the field of perception while preserving the information of small-scale,and improving the detection capability of large-scale individuals,making it better adapted to the multi-scale changes of the population.Experimental results on the three public population counting datasets show that the proposed model has further improved the counting accuracy,with an MAE of 51.21 and an MSE of 83.70 on the ShanghaiTech Part A dataset.

关 键 词：密集场景 人群计数 空洞卷积 动态特征选择 点预测 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]