【作者】 王帅；

【导师】 张向刚；

【作者基本信息】 电子科技大学 ， 系统工程， 2021， 硕士

【摘要】 近年来,步态识别等生物特征领域的研究越来越被人们所关注。步态识别不同于人脸、虹膜识别等静态识别,这种方法可以通过在远处拍摄视频来进行识别,而不需要使用过多局部细节,克服了当前生物识别的一些局限性。步态识别技术优势明显,可以克服当前人脸识别等方法的缺点,能够在复杂场景下获得较为广泛的应用。提高步态身份识别的准确率有助于提升效率、减少工作量,尤其在车站行人身份检测、身份验证等方面的效率将会有非常大的进步。在步态身份识别中有几大难点,其中一个目前亟待解决的难点是跨视角问题又称多视角问题。本文将针对这个问题,通过深度学习等方法开展研究,主要工作如下:(1)调研了目前国内外针对步态身份识别和多视角问题的科研现状,比较了不同研究方法的差别,分析了各类方法的优劣,总结了各类可以提升识别准确率的方法的关键因素,为本文研究的开展做了前期理论支撑。(2)基于生成对抗网络的视角转换模型研究。多视角问题的关键在于无法描述各个视角间的关系,无法通过一个视角来识别另一个视角。本文通过训练生成对抗网络,让网络可以将其它视角的步态序列转换到一个统一视角来进行识别。本文展示了将其它角度转换到统一视角36度的效果图,肉眼就可以看到视角转换有较好的效果,同时,使用普通的机器学习方法识别转换后的视角获得了将近80%的准确率,充分证明了视角转换模型对于解决多视角问题有较好的效果。(3)基于人体姿态的步态识别模型研究。步态识别必须使用一个连续的步态序列,同时使用步态中的动态信息和静态信息而不能只是使用几帧静态的图像。如何更好使用地动态信息是当前步态识别的一大难点。本论文使用了从人体姿态坐标提取的人体关节长度、关节角度、角度加速度等特征组成特征矩阵用于识别行人身份。还结合长短时记忆网络和卷积神经网络自动提取行人步态序列中的动态信息和静态信息。同时还通过结合不同损失函数让模型更加鲁棒和通用。最后通过CASIA-B数据集验证了本文步态识别模型的识别准确率能达到90%,尤其在跨视角时还能保持识别准确率在91%左右,证明了本文方法在跨视角方面的优越性。实验证明本文提出的视角转换模型和基于人体姿态的步态识别模型在解决多视角问题中有较为理想的效果。更多还原

【Abstract】 In recent years,more and more attention has been paid to the research of gait recognition and other biometric characteristics.Gait recognition is different from static recognition such as face and iris recognition,which can be done by shooting video at a distance.It does not need to use too many local details and overcomes some limitations of current biometrics.Gait recognition technology has obvious advantages,which can overcome the shortcomings of current face recognition methods.It can be widely used in complex scenarios.Improving the accuracy of gait identification is helpful to improve the efficiency and reduce the workload,especially in the station pedestrian identity detection,identity verification and other aspects of the efficiency will have a great progress.There are several difficulties in gait identification,one of which needs to be solved urgently is the cross-view problem,also known as the multi-view problem.In view of this problem,this paper will carry out research through deep learning and other methods.The main work is as follows:(1)The research status of gait identification and multi-perspective problems at home and abroad was investigated.The differences of different research methods were compared.The advantages and disadvantages of various methods are analyzed.The key factors of various methods to improve the accuracy of identification are summarized.It is a preliminary theoretical support for the development of this study.(2)Research on Perspective Shift Model Based on Generative Adversarial Network.The key to the multi-perspective problem is that it is impossible to describe the relationship between different perspectives,and it is impossible to identify one perspective from another perspective.In this paper,an adversarial network is generated by training so that the network can transform the gait sequence from other perspectives into a unified perspective for recognition.This paper shows the renderings of converting other angles to a unified Angle of view of 36 degrees.The visual Angle conversion can be seen with the naked eye with good effect.At the same time,the common machine learning method is used to identify the converted perspectives with an accuracy rate of nearly 80%,which fully proves that the perspective transformation model has a good effect on solving multi-perspective problems.(3)Research on gait recognition model based on human posture.Gait recognition must use a continuous sequence of gaits,using both dynamic and static information from the gait rather than just using a few frames of static images.How to make better use of dynamic information is a major difficulty in gait recognition.In this paper,features such as joint length,joint Angle and angular acceleration extracted from human posture coordinates are used to form an eigenmatrix to identify pedestrian identity.The dynamic and static information of pedestrian gait sequence is extracted automatically by combining short and long time memory network and convolutional neural network.At the same time,different loss functions are combined to make the model more robust and universal.Finally,the CASIA-B data set verified that the recognition accuracy of the gait recognition model in this paper can reach 90%,and the recognition accuracy can be maintained at about 91%in the case of cross-view,which proves the superiority of the proposed method in cross-view.The experimental results show that the perspective conversion model and the gait recognition model based on human body pose proposed in this paper have a relatively ideal effect in solving the problem of multi-perspective.更多还原