基于深度学习的步态识别关键技术研究
发布时间：2018-11-05 17:58
【摘要】：生物识别技术是计算机视觉领域一个前沿的研究课题。在众多的生物特征之中,步态具有可远程获取、鲁棒性强、安全性强等优势。因此,在“以人为中心”的现代智能监控系统中,步态识别技术获得了越来越多的关注。然而此问题存在着众多挑战,比如相同目标因不同视角、穿戴和行走速度带来的类内差异太大,以及不同目标之间的形态相似性带来的类间语义模糊等。目前的步态识别技术大多基于人工视觉特征来进行模型匹配,但是传统的人工特征已经无法满足步态精细识别的需求,所以很难打破特征提取和特征表示的瓶颈。在本文中,我们围绕基于深度学习的步态识别问题,提出了一系列新模型和新方法。首先,我们设计了一个基于深度学习的步态识别技术框架。为了克服现有步态数据库样本容量小以及深度学习训练速度慢的挑战,我们将原始的步态序列进行融合,计算其步态能量图作为卷积神经网络的输入来对预训练的网络进行微调。然后,我们提出了基于Siamese神经网络的步态识别技术。该技术借助深度神经网络的视觉特征学习能力与Siamese结构的距离度量学习特性,有效解决了深度学习训练数据量不足以及分类与识别任务的领域鸿沟问题。最后,我们通过联合步态序列的三维卷积特征和Siamese结构在三维空间进行特征度量学习。该方法可以从连续的周期性步态序列中捕捉空间维度和时间维度的信息,进一步提高步态识别的准确率和实用性。经实验验证,本文提出的方法在步态属性分类和身份识别中都取得了理想的结果,特别是在身份识别任务中,在目前世界上最大的步态数据库中,本文算法相比已有最好方法在正确识别率方面平均提高了5%。
[Abstract]:Biometrics is a frontier research topic in the field of computer vision. Among the many biological features, gait has the advantages of remote acquisition, robustness and security. Therefore, gait recognition technology has gained more and more attention in the modern intelligent monitoring system. However, there are many challenges to this problem, such as the same target is different from different angles of view, the intra-class differences caused by wearing and walking speed are too big, and the semantic ambiguity between classes caused by the morphological similarity between different targets and so on. Most of the current gait recognition techniques are based on artificial visual features for model matching, but the traditional artificial features can no longer meet the needs of fine gait recognition, so it is difficult to break the bottleneck of feature extraction and feature representation. In this paper, we propose a series of new models and methods for gait recognition based on deep learning. Firstly, we design a gait recognition framework based on deep learning. In order to overcome the challenge of small sample size and slow training speed in the existing gait database, we fuse the original gait sequences. The gait energy diagram is calculated as the input of the convolutional neural network to fine-tune the pretrained network. Then, we propose a gait recognition technique based on Siamese neural network. With the help of the visual feature learning ability of the deep neural network and the distance metric learning characteristic of the Siamese structure, this technique effectively solves the problem of insufficient data amount of in-depth learning training and the domain gap between classification and recognition tasks. Finally, we study the feature metrics in 3D space by combining the 3D convolution features of gait sequences and the Siamese structure. This method can capture the information of spatial dimension and time dimension from continuous periodic gait sequences and further improve the accuracy and practicability of gait recognition. Experimental results show that the proposed method has achieved ideal results in gait attribute classification and identification, especially in the task of identification and in the world's largest gait database. Compared with the best method, the algorithm improves the correct recognition rate by an average of 5%.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41;TP181
本文编号：2312872
[Abstract]:Biometrics is a frontier research topic in the field of computer vision. Among the many biological features, gait has the advantages of remote acquisition, robustness and security. Therefore, gait recognition technology has gained more and more attention in the modern intelligent monitoring system. However, there are many challenges to this problem, such as the same target is different from different angles of view, the intra-class differences caused by wearing and walking speed are too big, and the semantic ambiguity between classes caused by the morphological similarity between different targets and so on. Most of the current gait recognition techniques are based on artificial visual features for model matching, but the traditional artificial features can no longer meet the needs of fine gait recognition, so it is difficult to break the bottleneck of feature extraction and feature representation. In this paper, we propose a series of new models and methods for gait recognition based on deep learning. Firstly, we design a gait recognition framework based on deep learning. In order to overcome the challenge of small sample size and slow training speed in the existing gait database, we fuse the original gait sequences. The gait energy diagram is calculated as the input of the convolutional neural network to fine-tune the pretrained network. Then, we propose a gait recognition technique based on Siamese neural network. With the help of the visual feature learning ability of the deep neural network and the distance metric learning characteristic of the Siamese structure, this technique effectively solves the problem of insufficient data amount of in-depth learning training and the domain gap between classification and recognition tasks. Finally, we study the feature metrics in 3D space by combining the 3D convolution features of gait sequences and the Siamese structure. This method can capture the information of spatial dimension and time dimension from continuous periodic gait sequences and further improve the accuracy and practicability of gait recognition. Experimental results show that the proposed method has achieved ideal results in gait attribute classification and identification, especially in the task of identification and in the world's largest gait database. Compared with the best method, the algorithm improves the correct recognition rate by an average of 5%.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41;TP181
【引证文献】
相关硕士学位论文 前1条
1 陈春利;基于集成深度学习的雷达信号识别方法研究[D];西南交通大学;2018年
，本文编号：2312872
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2312872.html
