基于视觉注意力机制的人群密度估计方法  

CROWD DENSITY ESTIMATION METHOD BASED ON VISUAL ATTENTION MECHANISM

作　　者：朱利华[1] 朱玲玲[2] Zhu Lihua;Zhu Lingling(School of Software and Big Data,Changzhou College of Information Technology,Changzhou 213164,Jiangsu,China;School of Information Science and Technology,Nantong University,Nantong 226200,Jiangsu,China)

机构地区：[1]常州信息职业技术学院软件与大数据学院,江苏常州213164 [2]南通大学信息科学技术学院,江苏南通226200

出　　处：《计算机应用与软件》2022年第4期228-233,共6页Computer Applications and Software

基　　金：江苏省高校自然科学基金项目(18KJB480001);全国高等院校计算机基础教育研究会2018年重点立项课题(GZYZD2018014);2017年江苏省"青蓝工程"优秀团队。

摘　　要：针对人群图像中尺度变化大及现有密度估计网络泛化性能差的问题,提出一种基于视觉注意力机制的人群密度估计方法。该网络模型基于VGG-16网络,由空间注意力模块(Spatial Attention Module, SAM)、全局注意力模块(Global Attention Module, GAM)及融合模块(Fusion Module, FM)等几个模块组成,通过在网络不同层级中引入注意力机制来选择性地增强网络的功能,提高多尺度级联的有效性。设计一个弱监督学习来扩展提出的密度估计网络,使人群密度估计模型可以适应不同的场景和数据集。实验结果表明,该方法对不同尺度、不同场景下的人群密度图像都有很好的适用性和准确性,算法性能也优于现有的人群密度估计算法。Aiming at the problems of large-scale changes in crowd images and poor generalization of existing density estimation networks, this paper proposes the crowd density estimation method based on visual attention mechanism. Based on the VGG-16 network, the network model was consists of several modules, such as the Spatial Attention Module(SAM), the Global Attention Module(GAM) and the Fusion Module(FM). We introduced the attention mechanism in different levels of the network to selectively enhance the function of the network and to improve the effectiveness of multi-scale cascading. The weak supervised learning was designed to extend the proposed density estimation network, so that the crowd density estimation model could adapt to different scenarios and data sets. The experimental results show that the proposed method has good applicability and accuracy for crowd density images in different scales and scenarios, and its performance is prior to the existing algorithm.

关 键 词：人群密度估计 注意力机制 弱监督学习 密度图 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]