【作者】 项宗杰；

【导师】 刘允才；

【作者基本信息】 上海交通大学 ， 模式识别与智能信息系统， 2013， 博士

【摘要】 随着摄像机的普及和人工智能的发展，视频监控系统被越来越广泛地用于各种场合。由于单个摄像机视域是非常有限的，很多监控系统使用多个视域不重叠的摄像机以监控更广阔的区域，因此，在摄像机网络中对人体目标进行跟踪成为了监控领域的重要研究课题。单摄像机跟踪算法仅对单个视域内的人体目标进行跟踪，要在多个视域互不重叠的摄像机构建的网络中对人体目标进行跟踪，必须把同一个人在不同视域中的轨迹关联起来，这一关联的过程即跨视域识别。但由于人的外观在不同的视角、不同的光照强度下存在极大的差异，这些差异对跨视域识别提出了极大的挑战。针对这些挑战，本文给出解决方案。本文的创新点和贡献如下：（1）本文阐述了跨视域识别的三种不同的框架以及与之对应的评估方法，这三种框架分别对应于三种不同的情形：1）表观数据库不存在；2）表观数据库存在但不完整；3）表观数据库存在且完整。（2）本文提出了模糊空间颜色直方图，并把它作为表观特征进行跨视域识别。实验证明该特征在跨视域识别中的准确度能够媲美领域内公认的优秀特征，且计算效率显著优于后者，能够应用于实时场合。本文使用相关性度量作为表观相似度的度量方法。确定了表观特征和相似度的度量方法，无论表观数据库是否存在，系统都可进行最基本的跨视域识别。（3）本文提出了在无重叠视域摄像机网络中表观数据库的学习方法，并把学得的表观数据库用于跨视域识别。因为学得的表观数据库包含了每个人在不同视域、不同角度下的表观模型，因此把该表观数据库应用于跨视域识别，很好地解决了视角的变化和不同视域间光照的差异给跨视域识别带来的挑战。（4）本文综合了特征提取、表观数据库的学习和跨视域识别算法，构建了跨视域识别系统。该系统能够在无重叠视域摄像机网络中对人体目标进行跨视域识别。实验表明，该系统取得了较高的跨视域识别正确率。更多还原

【Abstract】 With the mass production of the cameras and development in artificialintelligence, video surveillance becomes popular in our daily life. As onecamera can only cover a limited field of view (FoV), multiple cameras aredeployed, composing a camera network with non-overlapped FoV. Trackinga person in such a camera network becomes an imperative task for the scien-tist in this area.Traditional tracking algorithm focuses on intra-camera tracking. Totrack people across the camera, person re-identification technique is needed.It is a process that associates different instances of the same person in differ-ent cameras and different times. It is critical, yet challenging, because theappearance of a person varies significantly in different viewpoints and dif-ferent light conditions. This dissertation provides the solutions to these chal-lenges. The following is the innovations and contributions of this disserta-tion: (1) This dissertation has stated clearly the re-identification frameworkand the evaluation method are different in three conditions:1) the gallery ofpersons is absent;2) the gallery is available but not complete;3) the full gal-lery is available.(2) This dissertation develops a feature, Fuzzy Space Color Histogram,for person re-identification purpose. The results of the experiments verifythat the performance of this feature is comparable to the state of art feature inperson re-identification, and with much lower computational cost. The over-all re-identification system can run in real time. The correlation is used as themetric of the appearance similarity. With the feature and similarity metric,the elementary re-identification system can be built, no matter whether thegallery is available or not.(3) This dissertation proposes the offline and online methods for learn-ing a gallery of persons who frequently appear in the camera network. Thegallery contains appearance models of these persons consisting of appearancefeatures of the person from each camera and viewpoint. Given the cameraidentity, viewpoint identity, person identity, the model is decided. Since theseappearance models are specific to each camera and viewpoint, the problemsof viewpoint variations and illumination variations between cameras are ex-plicitly solved. Experiments demonstrate that the framework provides signif- icant improvement in addressing the re-identification problem.(4) This dissertation presents a re-identification system that includes thefeature extraction module, the gallery learning module and there-identification module. Experiments demonstrate the high performance ofthis system in addressing the re-identification problem.更多还原