非重叠域行人再识别算法研究  

Research on pedestrian re-recognition in non-overlapping domain

作　　者：何晴 郭捷[1] HEQing;GUOJie(School of Electronic Information and Electrical Engineering,Shanghai Jiaotong University,Shanghai 200240,Chin)

机构地区：[1]上海交通大学电子信息与电气工程学院,上海200240

出　　处：《信息技术》2018年第7期34-38,共5页Information Technology

基　　金：国家重点研发项目(2017YFB1002401)

摘　　要：随着多摄像头监控系统的广泛应用,非重叠域中行人目标的再识别成为智能视频监控研究的热门领域。目前,非重叠域行人再识别的研究主要集中在两个方面:特征提取和度量学习。文中提出了一种基于深度卷积生成对抗网络(DCGAN)的行人再识别方法,该方法训练DCGAN网络来提取特征,DCGAN网络包括两个部分:生成模型G,判别模型D,训练过程中通过两个模型的对抗,可以得到行人图片更具区分性的特征描述,然后根据提取的特征进行相似性度量,判断是否是同一个行人。实验结果表明该方法对于非重叠域的行人再识别具有一定效果,而且可以消除清晰度、光照、遮挡等的负面影响,具有一定的鲁棒性。With the wide application of multi-camera monitoring system, the re-identification of pedestrian targets in non-overlapping domains has become a hot area of intelligent video surveillance research. Research on pedestrian re-identification mainly focuses on two aspects： feature extraction and measurement learning. In this paper,a method based on deep convolutional generative adversarial networks（DCGANs） is proposed. This method trains DCGAN network to extract features. DCGAN network consists of two parts： generator G, discriminator D. This method can get the more distinguishable feature description of the pedestrian pictures by the adversarial train,and then determine the picture whether it is the same pedestrian by similarity measurement according to the extracted feature.The results of the experiment show that this method has a certain effect on non-overlapping pedestrian rerecognition,and can eliminate the negative effects of sharpness,illumination and occlusion,and has certain robustness.

关 键 词：行人再识别 卷积神经网络 生成对抗网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]