摘要
针对密集场景下人群图像拍摄视角或距离多变造成的多尺度特征获取不足、融合不佳和全局特征利用不充分等问题,提出一种比例融合与多层规模感知的人群计数网络。首先采用骨干网络VGG16提取人群密度初始特征;其次,设计多层规模感知模块,获得人群多尺度信息的丰富表达;再次,提出比例融合策略,根据卷积层捕获的特征权重重构多尺度信息,提取显著性人群特征;最后,采用卷积回归策略进行密度图的回归。同时,提出一种局部一致性损失函数,通过区域化密度图的方式增强生成密度图与真实密度图的相似度,提高计数性能。在多个人群数据集上的试验结果表明,所提模型优于近年人群计数的先进方法,且在车辆计数上有较好推广性。
To deal with the problems of insufficient multiscale feature acquisition,poor fusion,and insufficient utilization of global features as a result of the changing view angles or distances of crowd images in dense scenes,we propose a crowd counting network based on proportion fusion and multilayer scale-aware.First,the backbone network VGG16 is employed to extract the initial characteristics of the population density.Subsequently,a multilayer scale-aware module is developed to acquire a rich expression of multiscale information from the crowd.Afterward,a proportional fusion strategy is designed to reconstruct the multiscale information based on the feature weights captured by the convolution layer and extract the significant crowd features.Lastly,convolution regression is utilized to regress the density map.Concurrently,a local consistency loss function is proposed,which improves the similarity between the generated density map and the real density map by regionalizing the density map and enhances the counting performance.The results of the experiments on multiple population datasets exhibit that the model proposed here surpasses the existing state-of-the-art methods of population density counting and has good generalization in vehicle counting.
作者
孟月波
张娅琳
王宙
MENG Yuebo;ZHANG Yalin;WANG Zhou(College of Information and Control Engineering,Xi’an University of Architecture and Technology,Xi’an 710055,China)
出处
《智能系统学报》
CSCD
北大核心
2024年第2期307-315,共9页
CAAI Transactions on Intelligent Systems
基金
陕西省重点研发计划项目(2021SF-429)。
关键词
人群密度估计与计数
卷积神经网络
多层规模感知
比例融合
局部一致性损失
密度图回归
多尺度信息
空洞卷积
crowd density estimation and counting
convolutional neural network
multilayer scale-aware
proportional fusion
local consistency loss
density map regression
multiscale information
dilated convolution
