优质期刊论文发表网提供专业的论文发表、论文写作以及期刊推广服务!QQ:820771224

电话:18796337551
当前位置:优质期刊论文发表网信息系统 → 文章正文

基于孪生网络的行人重识别研究

作者:优质期刊论文发表网  来源:www.yzqkw.com  发布时间:2019/10/10 9:31:15  

摘要:随着社会的快速发展和人们的安全意识逐渐提高,社会公共安全越来越受到关注。大量的摄像头被安装在诸如商场、火车站等公共场所,其采集的监控视频数据为公安部门侦破案件提供了线索,为交管部门管制交通提供了实时路况信息,还可以帮助搜寻失踪人口等,监控视频为人们的人身和财产安全带来了保障。而监控视频数据量在不断增大,人工查找目标的方式难以应对海量增长的视频,因此使用计算机视觉技术来高效率检索目标行人成为迫切的需求,行人重识别(Person Re-identification, Re-ID)的课题研究也由此展开。行人重识别是指在非重叠(不相交)视域多摄像头下的行人匹配。本文针对行人重识别的难点,使用深度学习方法来开展研究,提出了两种基于孪生网络的行人重识别方法,主要工作如下:

(1)提出一种基于孪生网络和重排序的行人重识别方法来处理行人重识别由于受到光照、行人姿势及遮挡等影响匹配精度和实验过程中存在图像错误匹配的问题。首先,给定一对行人训练图像,孪生网络可以同时学习一个具有辨别力的卷积神经网络(Convolutional Neural Network, CNN)特征和相似性度量,并预测两个输入图像的行人身份以及判断它们是否属于同一个行人,然后通过k互近邻方法来降低图像错误匹配的情况,最后将欧氏距离和杰卡德距离加权来对排序表进行重排序。

(2)提出一种基于全卷积孪生网络和注意力机制的行人重识别方法来解决行人重识别面临摄像头视角变化、行人距离摄像头远近及遮挡等挑战和实验中存在行人定位和错位的问题。具体来说,给定一对行人图像,首先使用三个卷积层来提取两个输入图像的特征,再将带有注意力机制的卷积相似性网络(Convolution Similarity Network, CSN)插入到不同的卷积层中以提取其视觉相似性,CSN可以改进两张图片相似性的计算;在识别模块中,首先将提取的特征向量传递给行人判别特征(ID-discriminative Embedding, IDE)分类器进行身份预测,然后从其身份标签中检索当前分支中输入图像的注意力映射,识别注意力损失再引导识别模块发现输入图像更完整的注意力区域,从而指导网络更好地定位感兴趣的行人。

(3)在行人重识别三大公开数据集Market-1501,CUHK03和DukeMTMC-reID上验证本文所提的两种基于孪生网络的行人重识别方法,实验结果显示,结合识别与验证模型的孪生网络有效地提高了行人重识别匹配率,加入重排序后匹配率得到进一步地提升,在数据集Market-1501上Single Query情况下Rank-1达到了83.44%,mAP达到了74.49%;结合注意力机制的全卷积孪生网络提取了更强大的特征,再次提高了精度,在数据集Market-1501上Single Query情况下Rank-1达到了89.20%,mAP达到了70.90%。所提方法均取得了较高的精度,超过了很多方法。

With the rapid development of society andthe gradual improvement of people's safety awareness, social public securityhas been paid more and more attention. Large number of cameras have beeninstalled in public places such as shopping malls and railway station, thesurveillance video data provides clues for the police to detect cases, providesreal-time traffic information for the traffic control department, and it canalso help search for missing persons and so on. Video surveillance providessecurity for people's personal and property safety. However, the volume ofsurveillance video data is constantly increasing, and it is difficult to copewith the massive increase of video by manual target search. Therefore, it hasbecome an urgent need to use computer vision technology to efficiently retrievetarget pedestrians, and the subject research of Person re-identification hasalso been carried out. Person re-identification refers to the matching ofpedestrians in non-overlapping (disjoint) vision field and multiple cameras.Aiming at the difficulty of person re-identification, this paper uses the deeplearning method to carry out research and proposes two siamese network-basedperson re-identification methods. The main work is as follows:

(1)Person Re-identification (Re-ID) undernon-overlapping multi-camera is easily affected by illumination,posture,and occlusion,and there are image mismatches in the experimental process.A Re-ID method based on siamese network and re-ranking was proposed.Firstly,a pair of pedestrian training imageswere given,a discriminative Convolutional NeuralNetwork (CNN) feature and similarity measure could be simultaneously learned bythe siamese network to predict the pedestrian identity of the two input imagesand determine whether they belonged to the same pedestrian. Then,the k-reciprocal neighbor method wasused to reduce the image mismatches. Finally,Euclidean distance and Jaccard distance were weighted to re-rank thesorted list.

(2) In view of the challenges in the fieldof Person Re-identification, such as the change of camera angle, the distancefrom the camera and the occlusion of the pedestrian, and the problems ofpositioning and dislocation of the pedestrian in the experiment, a personre-identification method based on the full convolutional siamese network andthe attention mechanism was proposed. In particular, given a pair of pedestrianimages, first, three convolutional layers were used to extract the features oftwo input images, then  ConvolutionSimilarity Network with attention mechanism (Convolution Similarity Network,CSN) was inserted into the different Convolutional layer in order to extractthe visual Similarity, CSN could improve the calculation of similarity betweentwo images; in recognition module, first the feature vector was extracted toIDE classifier for identity prediction, and then the attention map of the inputimage in the current branch was retrieved from its identity label, theidentification attention loss guided the recognition module to discover themore complete attention area of the input image, so as to guide the network tobetter locate the interested pedestrian.

(3) The two siamese network-based personre-identification methods proposed in this paper were verified on three majorpublic datasets of person re-identification: Market-1501, CUHK03 andDukeMTMC-reID. The experimental results showed that the siamese networkcombined with the identification and verification models effectively improvedthe matching rate of person re-identification, and the matching rate wasfurther improved after re-ranking was added. In the case of Single Query on thedataset Market-1501, the Rank 1 reached 83.44%, and the mAP reached 74.49%. Thefull convolutional siamese network combined with the attention mechanismextracted more powerful features, which again improved the accuracy.

关键词:行人重识别;孪生网络;重排序;卷积相似性网络;注意力

person re-identification; siamese network;re-ranking; convolution similarity network; attention

联系方式

客服QQ 820771224
客服热线18796337551
网站地址 www.yzqkw.com
郑重承诺 高效,快速,包发表!
优质期刊论文发表网真诚欢迎新老客户的光临与惠顾!