摘要
多视图人群计数,指从多个视角的图片中估计当前场景的人数.大多数基于卷积神经网络的方法使用多个同构但独立的分支处理不同视图,在增加模型复杂性的同时,引入大量的冗余特征.针对此问题,本文提出一种基于视图分支共享的卷积神经网络模型,使用同一视图分支从不同视角的图片提取出多个尺度对齐的特征图.这些特征图被投影到同一个世界平面上进行融合,进而回归出当前场景人群分布密度图.在视图分支内部,该模型在保持一定的结构复杂性的同时,减少各卷积层的核数量,极大降低模型可学习的参数数量.本文在两个公开数据集(PETS2009、CityStreet)上测试了性能,与5种已有方法相比较,本文方法能达到更好的性能.
Multi-view crowd counting aims to estimate the crowd count of current scene with images from multiple views.Most of multi-view crowd counting methods based on deep convolutional neural network have use independent view branches with the same structure, which increases the model′s complexity and brings massive redundant features.In this paper, we propose a new convolutional neural network model that based on sharing view branch.Our model uses the same view branch to extract several scale-consistent feature maps with images from different views.These feature maps will be projected into the same plane in world space to fuse, then regressed to a density map about crowd of current scene.In the view branch, we use less convolutional kernels under the structure with a certain complexity, which greatly decreases the number of learnable parameters.Extensive experiments are conducted on two opening datasets which are PETS2009 and CityStreet.Results establish the higher performance of our method over many state-of-the-art methods.
作者
王永会
涂可
郦洋
WANG Yong-hui;TU Ke;LI Yang(School of Computer Science and Engineering,Shenyang Jianzhu University,Shenyang 110168,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2023年第3期582-588,共7页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(62073227)资助
国家重点研发计划项目(2020YFC0833203,2021YFF0306303)资助。
关键词
人群计数
卷积神经网络
多视图
分支共享
crowd counting
convolutional neural network
multi-view
branch sharing
