基于主颜色谱及空间分布熵的目标匹配技术分析-analysis of target matching technology based on principal color spectrum and spatial distribution entropy.docxVIP

摘要由于智能交通、医疗、安防等各个领域的迫切需求，智能监控中的多摄像机协同关键 技术越来越受到人们的关注，其主要包括目标检测和定位、目标匹配、目标交接和目标行 为分析等技术。目标匹配技术作为实现多摄像机协同的桥梁近年来亦成为智能监控领域研 究的热点和难点。本文主要对目标匹配方法做了研究，首先介绍了常用的目标匹配方法，然后深入分析 了基于颜色特征的目标匹配方法，针对相关方法中存在计算量大、未考虑颜色空间分布信 息、目标匹配不够准确等问题，本文主要从以下几个方面进行了研究：(1) 提出了一种基于二级中心优化和 M-Kmeans 融合的目标主颜色谱描述算法。即在 用 Kmeans 聚类算法对目标的颜色进行聚类以降低计算复杂度时，先用二级中心优化算法 确定聚类初始中心点，然后用 M-Kmeans 算法进行聚类过程中的类成员调整。研究及实验 结果表明，融合算法不仅提高了目标主颜色谱描述的准确性，而且降低了聚类结果对初始 中心点的敏感性，提高了聚类结果的稳定性。(2) 提出了主颜色谱空间分布熵技术。即对于目标的主颜色谱，依据其对应像素在不 同划分区域的分布比例，应用香农熵公式计算主颜色谱分量的熵以表示其空间分部信息。 实验结果表明，通过主颜色谱空间分布熵能够很好地区分目标主颜色谱直方图相似但颜色 空间分布不同的目标。(3) 基 于 前 两 点 研 究 结 果 ， 提 出 了 MCS-SDE(Major Color Spectrum and Spatial Distribution Entropy)目标匹配方法。即用主颜色谱进行颜色相似度计算的同时，用该谱分 量的空间分布熵相似度模型进行加权,以此计算目标的匹配度。并在最后讨论了基于 MCS-SDE 目标匹配的多帧联合匹配方法。实验结果表明，新的目标匹配方法与文献[11] 的方法比较，提高了匹配对与非匹配对的区分度，匹配判定阈值更易确定，提高了目标匹 配的准确性和鲁棒性。本文最后对所研究的内容进行了总结，并对本课题可继续深入研究的方向进行了展 望。关 键 词：目标匹配、Kmeans 聚类、主颜色谱、空间分布熵AbstractDue to the urgent needs in many fields such as intelligent transportation, medical treatment and security, multi-camera monitoring techniques including object detection and location, object matching, object behavior analysis and so on, have been paid high attention. And object matching as the bridge of multi-camera monitoring has also become the keen and difficult topic in this field.The methods of object matching are the main work in this thesis. Firstly some common object matching methods are introduced and the methods based on the color characteristic are analised further more. And then, some problems of the methods based on color are pointed out, such as a large amount of calculation, insufficient color spatial distribution, inaccurate matching method and so on. The main contributions for the thesis are as follows:A major color spectrum description method based on fusion of the second class center optimizing and M-Kmeans is proposed. The second class center optimizing is used to choose the start points of the Kmeans clustering for obtaining color spectrum to reduce the amount of calcu