摘要
随着国家对社会公共安全的日益重视,无重叠视域监控系统已大规模的普及.行人再识别任务通过匹配不同视域摄像机下的行人目标,在当今环境下显得尤为重要.由于深度学习依赖大数据解决过拟合的特性,针对当前视频行人再识别数据量较小和学习特征单一的问题,我们提出了一种基于视频的改进行人再识别方法,该方法通过生成对抗网络去生成视频帧序列来增加样本数量和加入了行人关节点的特征信息去提升模型效率.实验结果表明,本文提出的改进方法可以有效地提高公开数据集的识别率,在PRID2011,iLIDS-VID数据集上进行实验,Rank 1分别达到了80.2%和66.3%.
As the government keeps attaching importance to public security,non-overlapping viewsheds surveillance systems have been deployed widely.It has become especially important to recognize pedestrian target through matching cameras with different viewsheds in nowadays.Deep learning relies on big data to solve overfitting.However,the current video-based person re-identification only has small data volume and homogeneous learning features.To solve this,we put forward a method to improve person re-identification based on the video.This method can increase the sample quantity by generating video frame sequence through generative adversarial network.It also adds the feature information of the pedestrian joints,which can improve the model efficiency.The experiment result shows that the modified method discussed in this paper can improve the recognition rate of public datasets effectively.In the experiments on PRID2011 and iLIDS-VID,Rank 1 attained 80.2%and 66.3%,respectively.
作者
刘一敏
蒋建国
齐美彬
刘皓
周华捷
LIU Yi-Min;JIANG Jian-Guo;QI Mei-Bin;LIU Hao;ZHOU Hua-Jie(School of Computer and Information,Hefei University of Technology,Hefei 230009;Anhui Province Key Laboratory of Industry Safety and Emergency Technology,Hefei 230009;YouTu Laboratory,Tencent,Hefei 230009)
出处
《自动化学报》
EI
CSCD
北大核心
2020年第3期576-584,共9页
Acta Automatica Sinica
基金
国家自然科学基金(61371155,61771180)
安徽省重点研究与开发项目(1704d0802183)资助。
关键词
行人再识别
深度学习
生成对抗网络
人体姿态估计
Person re-identification
deep learning
generative adversarial network(GAN)
human pose estimation
