基于回归模型的全卷积网络人群计数算法  

Full convolutional networks for crowd counting algorithms based on regression model

作　　者：吴晓燕 WU Xiao-yan(Intelligent Manufacturing College,Sichuan University of Arts and Science,Dazhou 635000,China)

机构地区：[1]四川文理学院智能制造学院,四川达州635000

出　　处：《计算机工程与设计》2020年第10期2867-2871,共5页Computer Engineering and Design

基　　金：四川省教育厅基金项目(18ZB0511)。

摘　　要：针对当前采用密度图回归方法估计人群数量时人数被高估的问题,提出一种基于计数回归和密度图估计的全卷积网络框架,采用计数回归与密度图回归相结合的方式对人群密度图进行估计。为训练模型参数和有效避免过拟合现象的出现,设计一种更深更轻且参数数量很少的完全卷积网络(full convolutional network,FCN)作为人群密度图估计器。实验结果表明,提出算法对密度分布不均和尺度不一的人群图像都有很好的适用性和准确性,算法性能优于现有的人群计数算法。Aiming at the current overestimation problem of the number of people when using the density map regression method to estimate the number of people,a full convolutional network framework based on count regression and density map estimation was proposed,in which the combination of counting regression and density map regression was used to calculate the population density map.To train model parameters and avoid over-fitting effectively,a deeper and lighter full convolutional network(FCN)with fewer parameters was designed as a population density map estimator.Experimental results show that the proposed algorithm has good applicability and accuracy for population images with uneven density distribution and different scales,and its performance is also better than the existing crowd counting algorithm.

关 键 词：全卷积网络 人群计数 计数回归 密度图回归 像素分类 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]