基于深度学习与属性学习相结合的行人再识别方法研究

发布时间：2018-04-04 05:28

本文选题：行人再识别　切入点：卷积神经网络　出处：《江苏大学》2017年硕士论文

【摘要】：行人再识别作为公共场所视频监控中最重要的技术之一,受到了研究者的广泛关注。目前,行人再识别方法普遍通过提取行人的颜色、纹理、形状等低层特征来进行行人的区分,而行人作为一种非刚性对象,这些人工设定的特征对于行人的判断并不是最好的,而且基于数值的低层特征缺乏语义表达能力,在行人再识别的实际应用中不易被用户所理解。此外,大多数行人再识别方法采取有监督的学习方式,依赖于大量带标签的训练数据,而在实际应用中,获取关于每个行人的大量带标签样本图像是不可能完成的任务。针对现有行人再识别中存在的这些问题,本文提出基于深度学习与属性学习相结合的行人再识别方法,主要内容如下:(1)提出基于无监督卷积神经网络与行人属性的行人再识别方法。该方法通过结合卷积神经网络的模型结构和卷积自动编码器的学习原理,无监督地对行人图像进行特征提取,避免了对带标签训练样本的依赖,同时通过这种数据驱动的特征学习方式获得更具有代表性的行人特征,从而提高行人再识别的准确率。在行人特征与行人类别间加入属性层,通过对行人图像的属性判断间接地进行行人类别的判断,赋予了行人再识别方法更好的语义表达能力和实用价值。在VIPe R行人数据集上的实验结果表明,与现阶段所提方法相比,该方法能有效解决行人再识别中对带标签数据的依赖问题和缺乏语义表达能力的问题,并有效提高了属性分类器的准确率。(2)提出基于无监督卷积神经网络与层次属性的行人再识别方法。该方法将行人图像按身体部位划分为互相重叠的若干分块,对每个分块针对性地提取特征并分配属性分类器,有效降低了冗余信息对分类器造成的干扰,进一步提高了属性分类器准确率。引入层次属性,利用粗、细粒度属性来对行人进行区分,使得行人再识别方法更加符合人们的认知规律,并能够应对不同程度行人描述时的再识别任务。在VIPe R行人数据集上,从多个方面验证了所提方法的有效性,实验结果表明该方法所取得的行人再识别准确率高于现有其他算法,并且对于属性缺失具有一定的容忍度。(3)设计并实现基于深度学习与属性学习相结合的行人再识别原型系统。采用MATLAB实现了行人再识别系统的开发并设计了简洁的GUI界面。系统包括有目标行人图像的再识别和无目标行人图像的再识别两大功能模块,主要包含目标行人图像输入、行人层次属性选择、行人再识别和候选行人图像展示等功能,验证了本文所提行人再识别方法的可用性。
[Abstract]:Pedestrian recognition as one of the most important public places in video surveillance technology, has attracted much attention of researchers. At present, the pedestrian recognition method by extracting common pedestrian color, texture, shape and other features to distinguish the low layer of pedestrians, and as a kind of non rigid object, the artificial set of features for the and determine the pedestrian is not the best, but the lack of low level features numerical semantic expression ability based on the practical application in the recognition of pedestrians are not easy to be understood by users. In addition, most of the pedestrian recognition method adopts supervised learning method, relies on a large number of training data with the label, but in practical application. For each pedestrian plenty of labeled sample image is an impossible task. According to the existing pedestrian recognition of these problems, this paper based on deep learning and attributes Recognition method of pedestrian combination of study, the main contents are as follows: (1) proposed unsupervised convolutional neural network and pedestrian pedestrian recognition method based on attribute. This method by learning principle combined with convolutional neural network model structure and automatic convolution encoder, the feature extraction of pedestrian images without supervision, to avoid the dependence on the labeled training samples, and feature driven by this data learning way of obtaining pedestrian characteristics more representative, so as to improve the accuracy of pedestrian recognition. The pedestrian characteristics and pedestrian categories added attribute layer, judge indirectly by category judgment through the attribute of the pedestrian pedestrian image, expression ability and practical the semantic value gives better. Then pedestrian recognition method in VIPe R pedestrian dataset. The experimental results show that, compared with the stage of the proposed method, this method can effectively To solve the problem of dependence on labeled data and the lack of semantic expression ability of re recognition of pedestrians, and effectively improve the accuracy of attribute classifiers. (2) proposed unsupervised convolutional neural networks with different attributes and pedestrian recognition method based on this method. The pedestrian image according to parts of the body are divided into several overlapping points for each block, the block to extract features and distribution of attribute classifier, effectively reduce the interference of redundant information to the classifier caused, to further improve the classifier accuracy. The attribute level attribute, using coarse and fine granularity attribute to distinguish pedestrian, pedestrian recognition method makes more in line with the cognition of the people, and to cope with the different degree of pedestrian description recognition task. In the VIPe R pedestrian dataset, the effectiveness of the proposed method are verified from many aspects, the experimental results show that the The method of pedestrian recognition accuracy than other existing algorithms, and for the missing attribute has a certain tolerance. (3) the design and implementation of pedestrian deep learning and attribute based learning combined recognition prototype system. Using the MATLAB to realize the development of pedestrian recognition system and design a simple GUI interface. No target recognition and image recognition of pedestrians two major functional modules of the system include target pedestrian images, including pedestrian target image input, the pedestrian level attribute selection, pedestrian recognition and candidate pedestrian image display and other functions, to verify the availability of recognition method this paper submitted for people.

【学位授予单位】：江苏大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】