【作者】 李强；

【导师】 田玉敏； 郭向英；

【作者基本信息】 西安电子科技大学 ， 工程硕士（专业学位）， 2019， 硕士

【摘要】 随着当今时代的发展,社会安全问题越来越受到公众的重视。城市智能监控作为防治公共犯罪的重要手段,在公共场所秩序管理、大型案件侦破以及嫌疑人追踪等方面的作用日益凸显。作为智能安全领域的主要研究方向之一,行人重识别(Person Reidentification)是一种利用计算机视觉算法在给定的图像集或视频序列中查找特定行人的技术。然而由于实际应用场景较为复杂,导致难以提取行人具有辨别力的有效特征,这给行人重识别任务带来了极大挑战。此外,传统检索方法无法满足海量视频数据任务需求的问题也逐渐浮现,因此如何准确且快速地定位目标行人成为亟待解决的问题。深度学习算法可以自适应地学习图像的有效特征,避免了人工设计和选择特征的弊端,并且可以适应大数据时代的数据量需求。因此,本文使用基于深度学习的方法,从以下几个方面对行人重识别技术进行研究。(1)针对大部分相关数据集中行人图像数量不足的问题,本文提出了一种基于生成式对抗网络的数据增强模型PersonGAN。该模型以CycleGAN为基础,并对网络进行相应改进,以适应行人重识别任务的需求。同时,本文设计了一种恒等映射损失(identify mapping loss)约束网络模型,从而可生成不同摄像头风格的行人图像。PersonGAN不仅可以有效地扩充数据集,而且能够生成一些识别难度更大的样本,给训练提供新的动量,使模型具有更强的鲁棒性和更广泛的适用性。(2)针对难以提取行人具有辨别力特征的问题,本文提出了一种鲁棒联合学习网络(Robust Joint Learning Networks,RJLN)。RJLN是基于表征学习的网络模型,首先通过基础网络提取行人的低级特征,然后使用改进的全局分支网络和改进的局部分支网络分别提取目标行人的全局特征和局部特征,以适应提取不同粒度特征的需求,提高模型的学习能力,从而使模型的识别准确率和鲁棒性得到显著提升。最后,融合两个分支网络的损失函数,作为RJLN的总体损失函数。(3)针对局部特征的提取问题,本文提出了一种结构简洁、性能优良的强化局部基础网络(Strong Part-base Networks,SP_Net)。与利用骨架关键点定位、姿态估计等局部特征提取方法不同,该模型直接以特征图水平分割后的条纹作为网络输入,之后通过全局平均池化从水平条纹中学习细粒度特征。为了提升模型的识别准确率和鲁棒性,本文在SP_Net中添加核尺寸为1×1的卷积层进行降维,并通过随机失活层和批正则化层来防止过拟合。通过在Market-1501、DukeMTMC-reID、CUHK03等数据集上进行实验,本文对模型性能的提升进行了验证。实验结果证明,改进后的算法在三个数据集上的rank-k准确率和mAP均有明显提高,与目前基于深度学习的行人重识别其他算法相比也具有一定的优越性。更多还原

【Abstract】 With the development of the current times,more and more public have paid attention to the social security.As a significant means of preventing and controlling public crimes,urban intelligent monitoring is playing an increasingly prominent role in public place management,large-scale case detection and suspect tracking.As one of the main research directions in the field of intelligent security,Person re-identification is a technology that using computer vision algorithms to find specific pedestrians in a given image set or video sequence.However,due to the complexity of the practical application scenario,it is difficult to extract the effective features of pedestrians with discrimination,which brings great challenges to the task of person re-identification.In addition,the problem that the traditional retrieval method can not meet the task requirements of massive video data is gradually emerging.As a result,how to locate the target pedestrian accurately and quickly becomes an urgent problem to be solved.The deep learning methods can adaptively learn the effective features of the image,so avoiding the disadvantages of manual design and selection of features,and can adapt to the data volume requirements of the big data times.Therefore,this thesis uses the method based on deep learning to study the person re-identification technology from the following aspects:First,for the problem that the images in most pedestrian datasets is insufficient,this thesis proposes a data augmentation model named Person GAN which based on the generative adversarial networks.The model is based on Cycle GAN and the network is modified to adopt to the needs of pedestrian recognition tasks.At the same time,this thesis designs an identity mapping loss to constraint network model,so as to generate pedestrian images with different camera styles.Person GAN can not only effectively expand the dataset,but also can generate some samples that are more difficult to identify,so that providing a new momentum for training,and making the model more robust and widely applicable.:Second,in view of the problem that it is difficult to extract the discriminative features of pedestrians,this thesis proposes a Robust Joint Learning Networks(RJLN).RJLN is a network model based on representation learning.Firstly,the low-level features of pedestrians are extracted through the basic network.Then the improved global branch network and the improved local branch network are used to extract the global features and local features of the target pedestrians respectively,so as to adapt to extract different granularity features and enhance the learning ability of the model,so that the identification accuracy and robustness of the model are significantly improved.Finally,the loss function of the two branch networks is merged as the overall loss function of RJLN.Third,aiming at the extraction of local features,a Strong Part-base Networks(SP_Net)with simple structure and excellent performance is proposed in this thesis.Different from the local feature extraction methods such as skeleton key point location and pose estimation,The model directly uses the horizontally partition stripes of the person feature map as the network input,and then learns the fine-grained features from the horizontal stripes through the global average pooling.In order to improve the identification accuracy and robustness of the model,a convolutional layer with a kernel size of 1×1 is added to the SP_Net for reducing the dimension,and the over-fitting is prevented by the dropout layer and the batch normalize layer.Through experiments on Market-1501,Duke MTMC-re ID,CUHK03 and other datasets,this thesis verifies the improvement of the model.The experimental results show that the improved methods has significantly higher rank-k accuracy and m AP on the three data sets,and outperforms existing deep learning methods.更多还原