中国科学院机构知识库网格
Chinese Academy of Sciences Institutional Repositories Grid
热门
跨摄像机行人识别与轨迹挖掘

文献类型:学位论文

作者胡杨
学位类别工学博士
答辩日期2015-05-27
授予单位中国科学院大学
授予地点中国科学院自动化研究所
导师李子青
关键词行人再辨识 特征设计 度量学习 深度学习 时空信息 轨迹挖掘 Person Re-identification Feature Design Metric Learning Deep Learning Spatio-temporal Information Trajectory Mining
其他题名Cross-View Person Recognition and Trajectory Mining
学位专业模式识别与智能系统
中文摘要近年来视频监控在安防领域得到了广泛应用,与之紧密相关的智能视频分析技术的研究成果也大量出现。然而现有的大多数研究都集中在单一摄像机领域。虽然利用多摄像机获取大范围视频信息的手段在实际应用中已经很普遍,但是研究领域相对滞后,缺乏有效的算法关联多个摄像机之间的信息。传统的高度依赖人工操作的多个视频处理方式在处理海量视频信息时效率极低,而多摄像机信息的融合能够为安防监控提供很多帮助。因此,计算机视觉领域新兴的跨摄像机行人识别(行人再辨识)引起了越来越多研究者的关注。为了进一步满足安防监控的实际应用需求,我们提出了轨迹挖掘的概念,旨在从宏观角度上解决感兴趣目标在多摄像机网络中的轨迹描述问题。行人再辨识是个困难的任务,仍有很多问题有待解决,例如跨摄像机造成的光照、视角、姿态和遮挡等变化,海量数据标注问题以及模型的通用性等。作为新概念的轨迹挖掘则有很多方面待完善,例如数据库收集、问题建模方式和轨迹评估方式等。 本文针对监控场景下的跨摄像机行人识别进行了深入研究,包括设计性能优异的行人特征表达、由训练集学习适合的距离测度、使用深度学习方法同时解决特征和分类器设计以及跨数据库测试等问题。行人再辨识是轨迹挖掘的重要组成部分,讨论完这个问题后,接着针对轨迹挖掘做了大量工作,包括概念的讨论和明确、轨迹挖掘数据库采集、基于表象和时间信息的轨迹挖掘方法和输出结果测试协议等。本文的研究工作和贡献主要包括如下内容: 1)提出了基于结构信息的多特征融合行人再辨识方法SCEFA,涉及到的三个特征在提取或者匹配时考虑行人的结构信息以提升特征的分辨能力和减少背景信息的干扰。加入的行人结构信息按照由小及大的顺序分为几个层次:像素级(pixel-level),斑点区域级(blob-level)和有语义的部件级(part-level)。SCEFA方法侧重于特征的提取及其对应的匹配方法,在不使用训练数据的情况下取得了很好的性能。该方法对应用场景的依赖性较低,有较好的通用性。提出了基于局部最大出现模式的特征提取与度量学习的行人再辨识方法。该方法中的局部最大出现模式(LOMO)特征设计方法对于跨摄像机造成的光照变化和视角变化做了针对性处理,而且仅仅使用了简单的SILTP 和HSV空间颜色直方图特征,具有很快的计算速度。然后同时学习最优子空间和该空间下的距离度量(XQDA),从而得到了更适合计算特征相似度的距离测度。该方法在多个常用行人再辨识数据库上对比了较多现有的方法。大量的实验结果表明该方法的特征设计和度量学习都非常高效,总体性能超过了现有的最好结果,特别是rank-1的识别率提升幅度较大。 2)提出了之前未受到足够重视的跨数据库行人再辨识问题,并通过使用卷积神经网络学习行人的特征表达的方式解决该问题。详细阐述了所使用的网络结构、参数设置和训练过程。多个测试集上的实验结果表明本文的方法学习到的行人特征表达具有很好的通用性,也为行人再辨识领域的行人特征表达问题提供了一个具有很好性能的通用解决方法。另一方面,通过对使用不同训练集学习到的行人特征表达进行性能评估,对使用...
英文摘要In recent years video surveillance technique is popularly used, related intelligent video analysis algorithms have been proposed in the literature. However, most researches focus on the single camera setting. Though multi-camera system is widely used nowadays for obtaining large field information, few up-to-date researches have been proposed to associate information between cameras. Traditional multi-camera processing method highly depend on manual work which is inefficiency when process huge number of videos. Fusion the information of multiple cameras can help the security surveillance a lot. Therefore, person re-identification attract more and more attention in recent years. In order to further satisfy the practical need of security surveillance, we propose the trajectory mining concept which aims to describe the trajectory of interested target in a multi-camera network from a macroscopic perspective. Person re-identification is a difficult task, there are a lot of work need to do, such as problems caused by complex illumination, viewpoint, pose and occlusion. In addition, the collection of data labeling and the generalization ability of the model are also need to be further studied. As a new concept, there are many imperfect aspects, such as dataset collection, problem formation and evaluation strategy. This thesis focuses on person re-identification in security surveillance, addressing many basic problems in image video analysis and pattern recognition, including design good person representation, learn proper distance metric from the training set, apply CNN to address the feature design and classification at the same time, evaluate the performance across dataset. Person re-identification is an important element of trajectory mining. After solving the person re-identification task, we also do lots of work for trajectory mining, such as the discussing of the concept, dataset collection, method based on appearance and temporal information and the performance evaluation. The main works and contributions of this thesis are as follows. 1) The thesis proposes a person re-identification method called SCEFA which based on exploring the structural information and fusing multiple feature. For improving the discrimination ability and reducing the affect of background, structural constraints are considered during the feature extraction step or feature matching step for the three involved features. The structural constraints can be divided into three levels: ...
语种中文
其他标识符201118014628040
源URL[http://ir.ia.ac.cn/handle/173211/6709]  
专题毕业生_博士学位论文
推荐引用方式
GB/T 7714
胡杨. 跨摄像机行人识别与轨迹挖掘[D]. 中国科学院自动化研究所. 中国科学院大学. 2015.

入库方式: OAI收割

来源:自动化研究所

浏览0
下载0
收藏0
其他版本

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。