无重叠视域多摄像机目标关联研究

发布时间：2018-03-10 23:19

本文选题：无重叠视域多摄像机目标关联　切入点：目标再识别　出处：《合肥工业大学》2014年硕士论文　论文类型：学位论文

【摘要】：随着广域视频监控的普及,多摄像机智能监控已经成为计算机视觉领域的研究热点。考虑到成本效益和公众隐私,大型视频监控网络存在大量的监控盲区,造成多摄像机之间目标跟踪的困难。无重叠视域多摄像机监控系统下的目标关联,又称目标再识别,主要用于消除摄像机之间的视觉差异和监控盲区的时空不确定性对目标跟踪的影响。目前,目标关联已经成为解决无重叠视域监控场景下目标跟踪的关键问题。本文对无重叠视域多摄像机目标关联进行了深入的研究,论文的主要工作和创新点如下：(1)提出了基于目标“特征差”矢量的非线性排序(NRFD)目标再识别算法。针对目标在不同监控系统的视觉差异对目标再识别的影响,与以往旨在消除这类影响的目标再识别算法不同,本文利用这些差异,也就是目标特征差,构造二分类模型,并通过RBF核SVM对候选目标进行排序,从而实现目标再识别。此外,针对负样本远多于正样本而造成的样本失衡问题,本文构建了基于巴氏系数的AP聚类预处理方法来选择典型的负样本,从而克服该问题造成的影响。在标准数据库上的实验结果表明,该算法的识别率优于相关方法。(2)考虑到基于信息衡量的JENSEN-SHANNON(JS)核函数能处理光照变化引起的颜色直方图偏移问题,本文提出了JS核判别分析(JSKDA)的目标关联算法。首先通过JS核函数把样本映射到高维空间,充分提取目标的可鉴别信息,然后在该空间应用改进的局部Fisher判别分析,抵御离群样本对目标关联的影响,从而降低关联错误率。(3)应用D-S证据理论融合监控网络的时空约束信息和NRFD、JSKDA的匹配信息,避免了模型之间的冲突,提高了目标关联的准确性。
[Abstract]:With the popularity of wide-area video surveillance, multi-camera intelligent surveillance has become a research hotspot in the field of computer vision. Considering the cost effectiveness and public privacy, large video surveillance network has a large number of blind areas. It is difficult to track targets between multiple cameras. The target association under the multi-camera surveillance system without overlap is also called target re-recognition. It is mainly used to eliminate the visual difference between cameras and the influence of space-time uncertainty in monitoring blind area on target tracking. Target association has become the key problem to solve the problem of target tracking in the scene of non-overlapping visual surveillance. The main work and innovation of this paper are as follows: (1) A nonlinear sorting NRFD-based target rerecognition algorithm based on "eigenvalue" vector is proposed, aiming at the influence of the visual difference of target in different monitoring systems on target re-recognition. In this paper, we use these differences, that is, target feature difference, to construct a two-classification model, and sort candidate targets by RBF kernel SVM, so as to realize target re-recognition. In order to solve the problem of sample imbalance caused by far more negative samples than positive samples, an AP clustering preprocessing method based on pasteurian coefficient is proposed to select typical negative samples. The experimental results on the standard database show that the recognition rate of the algorithm is better than that of the related method. 2) considering that the JENSEN-SHANNON JSS kernel function based on information measurement can deal with the color histogram offset caused by the illumination change, In this paper, the target association algorithm of JS kernel discriminant analysis (JS kernel discriminant analysis) is proposed. Firstly, the sample is mapped to high dimensional space by JS kernel function, and then the discriminable information of the target is fully extracted, and then the improved local Fisher discriminant analysis is applied in this space. To resist the influence of outlier samples on target association and reduce the correlation error rate, D-S evidence theory is applied to fuse the spatio-temporal constraint information of monitoring network and the matching information of NRFD-JSKDA to avoid the conflict between models and improve the accuracy of target association.
【学位授予单位】：合肥工业大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN948.41

【相似文献】