【作者】 王宇；

【导师】 夏懿； 许胜强；

【作者基本信息】 安徽大学 ， 控制工程（专业学位）， 2022， 硕士

【摘要】 如今,人类社会正在朝着数字化和信息化的方向发展,个人与公共信息安全也遇到了前所未有的挑战,在这样的大环境下,生物识别技术的发展为保障信息安全与维护社会稳定提供了重要的技术支撑。迄今为止,指纹识别、虹膜识别和人脸识别仍然是人类社会中最普遍、最重要的生物识别方法,这得益于它们可以提供非常高的识别精度。然而,为了捕获这些方法所需要的生物特征,要求被识别对象主动配合,且这个过程通常只能在近距离实现。相比之下,步态这一生物特征可以以更隐蔽的方式远距离捕获且具有不易伪装和隐藏的特点。由于这些独特的优势,步态识别近年来备受关注。然而,步态识别技术的实际应用仍然面临诸多挑战,例如步行速度、观测视角、衣着和携带物体都会在不同程度上影响识别精度。这其中,观测视角的变化影响最大,因此如何有效解决跨视角步态识别问题已成为该领域的研究重点。本文基于深度学习技术开展跨视角步态识别问题研究,提出了两个不同的基于对抗神经网络的跨视角步态识别模型,并设计了一个端到端的跨视角步态识别软件系统,具体工作如下:(1)调研有关跨视角步态识别的国内外研究现状,并整理跨视角步态识别的发展历程,同时分析讨论各类方法的原理并总结其优势与局限性,从而为本文方法的提出进行理论铺垫。(2)步态能量图像(Gait Engery Image,GEI)有效融合了步态轮廓图序列中的静态和动态信息,可以大幅降低步态样本的数据大小,同时还降低了轮廓图中噪声的干扰,近年来被广泛应用于步态时空特征分析。本研究也采用步态能量图作为模型的输入数据。为此,本文基于PSP-Net-50(Pyramid Scene Parsing Network)网络并配合注意力机制(Squeeze-and-Excitation Networks,SENet)搭建了一个针对步态视频图像的步态轮廓分割网络,旨在将实时步态图像转换为轮廓图,进而得到步态轮廓图像序列的平均——步态能量图像。(3)提出了两种基于GAN的步态识别网络:首先提出了双通道循环一致性生成对抗网络(Two-channel Cycle Consistency Generative Adversarial Network,TCC-GAN),该网络可以利用单个模型来实现跨任意视角的步态图像转换。在该模型中,与视角相关的信息被视为可以在不同步态图像之间传递的属性;同时设计了一个受三重损失约束的身份保持模块,以保证步态图像在视角转换过程中的身份信息得以保存;此外本研究在模型的损失函数中引入循环一致性损失(Cycle Consistency Loss),用于缓解模式崩塌问题。在步态数据集CASIA-B和OU-MVLP上的实验证明,TCC-GAN模型的跨视角识别精度高于目前大多数其他基于GAN的跨视角步态识别模型。为了克服循环一致性损失所带来的隐藏噪声问题,本文进一步提出特征分布一致性生成对抗网络(Feature Distribution Consistent Generative Adversarial Network,FDC-GAN),该模型在TCC-GAN的基础上做出改进,提出采用周期性循环一致损失(Recurrent Cycle Consistency Loss)来去除生成图像中不易察觉的无用信息,进一步弥合生成图像的特征分布与真实分布之间的差距。实验证明,FDC-GAN在步态数据集CASIA-B和OU-MVLP上的识别精度相较TCC-GAN有着进一步的提高。(4)基于上述步态轮廓提取模型以及跨视角步态识别模型,设计并实现了一个基于视频的步态识别软件系统,该系统可以实现端到端的基于实时步态视频输入的身份识别,并同时具有相应的人员信息管理功能。更多还原

【Abstract】 Today,human society is developing towards digitization and informatization.In the meantime,personal and public information security has also encountered unprecedented challenges.In such a large environment,the development of biometric technology provides important technical support for ensuring information security and maintaining social stability.So far,fingerprint recognition,iris recognition and face recognition are still the most common and important biometric methods in human society due to the very high recognition accuracy they can provide.However,in order to capture the biometrics required by these methods,active cooperation of the identified object is required,and this process is usually only possible at close distance.In contrast,gait,a biological feature that can be captured at a distance in a more stealthy manner,is not easily camouflaged and altered.Due to these unique advantages,gait recognition has received much attention in recent years.However,the practical application of gait recognition technology still faces ma ny challenges,such as walking speed,view,clothing and carrying objects.These factors will affect the recognition accuracy to varying degrees.Among them,the change of view has the greatest impact,so how to effectively solve the problem of cross-view gait recognition has become the focus of research in this field.In this paper,we carry out research on cross-view gait recognition based on deep learning technology,propose two different cross-view gait recognition models based on adversarial neural networks,and design an end-to-end cross-view gait recognition software system.The specific work is as follows :(1)A thorough investigation on the research status of cross-view gait recognition at home and abroad is performed.We reviewed the technique development of this research field,analyzed and discussed the theoretical principles of various methods,and summarized their advantages and limitations.By this process,the theoretical foundation for the proposed method is setup.(2)The Gait Energy Image(GEI)effectively integrates the static and dynamic information in the gait silhouette sequences,which can greatly reduce the data size of the gait sample and the interference of noise in the silhouette image.It is widely used in the analysis of gait spatiotemporal feature.This study also adopted the GEI as the input data of the model.To this end,this paper builds a gait silhouette segmentation network with its inputs being gait video images based on the PSP-Net-50(Pyramid Scene Parsing Network)network and the attention mechanism(Squeeze-and-Excitation Networks,SENet).The real-time gait image is converted into a gait silhouette image,and then the average of the sequence of gait silhouette images,i.e.,the gait energy image,is obtained.(3)Two GAN-based gait recognition networks are proposed: First,a Two-channel Cycle Consistency Generative Adversarial Network(TCC-GAN)is proposed,which can utilize a single model to achieve gait image transformation from arbitrary views.In this model,the information related to the view is regarded as an attribute that can be transferred between different images.Meanwhile,an identity preservation module constrained by triple loss is designed to ensure the preservation of identity information of gait images in the process of view conversion.In addition,this study introduces Cycle Consistency Loss in the loss function of the model to alleviate the problem of mode collapse.Experiments on the gait datasets CASIAB and OU-MVLP demonstrate that the cross-view recognition accuracy of the TCC-GAN model is higher than most other current GAN-based cross-view gait recognition models.In order to overcome the hidden noise problem caused by the loss of cycle consistency,this paper further proposes a Feature Distribution Consistent Generative Adversarial Network(FDCGAN),which is improved on the basis of TCC-GAN.The Recurrent Cycle Consistency Loss is proposed to remove the useless information that is not easily perceptible in the generated image,and further bridge the gap between the feature distribution of the generated image and the real distribution.Experiments show that the recognition accuracy of FDC-GAN on gait datasets CASIA-B and OU-MVLP is further improved compared to TCC-GAN.(4)Based on the above gait silhouette segmentation model and cross-view gait recognition model,a video-based gait recognition software system is designed and implemented,which can realize end-to-end real-time gait-video-based identity recognition,and has corresponding module to realize personnel information management.更多还原