基于最大池图匹配的形变目标跟踪方法

发布时间：2018-07-23 15:19

【摘要】：随着大数据时代的到来,计算机技术和网络技术突飞猛进的发展,计算机视觉技术成为信息科学研究领域的重要课题。而作为诸多计算机视觉高层应用的基础,视觉跟踪技术也越来越受到国内外研究者的重视。根据实际应用,视觉跟踪主要分为两个大的方向：单目标跟踪和多目标跟踪。虽然研究者在单目标跟踪课题上做了大量的研究,但是目标在运动过程中所包含的各种信息以及场景限制并未得到充分挖掘。单目标跟踪过程中,目标可能会产生巨大形变或者面临严重遮挡,此时目标的外观会发生巨大变化,这种情况下如果继续使用传统的整体框(bounding box)来描述目标,势必会滤掉前景目标部分或者引入背景噪声,无法给出精确的目标表达。本文针对单目标跟踪进行相应研究和探讨,就跟踪过程中出现的关键技术难题,提出了基于部件的最大池图匹配的跟踪方法(Max-pooling Graph matching based Tracker, MGT)。文章的主要内容总结如下：(1)不同于基于目标整体模型的算法,本文算法基于目标部件模型,采用动态图结构表示目标部件,即目标部件的表象特征(表象信息),以及它们之间的相对位置关系(结构信息)。对于目标搜索区域,算法基于图像分割技术提取出超像素候选目标部件建立候选图,并与建立好的的目标图模型进行匹配。(2)图匹配策略采用最大池(max-pooling)图匹配方法,即目标图匹配对中的每一个节点支持项都只使用候选图中的最大池支持项,并将其相关结构一致性分数作为匹配似然度,建立起目标图模型和候选图之间的部件匹配关系。在此基础上得到目标位置的置信图(confidence map),通过采样可以确定目标的最优位置。(3)最后,为了避免仅考虑局部目标部件的贡献造成的鉴别力不够,我们引入了整体目标的特征表达参与目标位置投票,以提高跟踪鲁棒性。
[Abstract]:With the coming of big data era and the rapid development of computer technology and network technology, computer vision technology has become an important subject in the field of information science research. As the foundation of many high-level applications of computer vision, visual tracking technology has been paid more and more attention by researchers at home and abroad. According to the practical application, visual tracking is divided into two major directions: single target tracking and multi-target tracking. Although researchers have done a lot of research on the subject of single target tracking, all kinds of information and scene constraints contained in the process of target motion have not been fully exploited. In the process of single target tracking, the target may produce huge deformation or face severe occlusion, and the appearance of the target will change greatly. In this case, if we continue to use the traditional global box (bounding box) to describe the target, It will filter out the foreground target or introduce background noise, so it can not express the target accurately. Based on the research and discussion of single target tracking, this paper presents a tracking method based on component maximum pool map matching (Max-pooling Graph matching based Tracker, MGT).) for the key technical problems in the tracking process. The main contents of this paper are summarized as follows: (1) different from the algorithm based on the whole object model, this algorithm is based on the target component model and uses dynamic graph structure to represent the target component. That is the representation of the target component (representation information) and the relative position relationship between them (structure information). For the target search region, the algorithm extracts candidate target components to build candidate images based on image segmentation technology, and matches them with the established target graph model. (2) the maximum pool (max-pooling) graph matching method is used to match the graph matching strategy. In other words, each node support item in the target graph matching pair only uses the maximum pool support item in the candidate graph, and takes the correlation structure consistency score as the matching likelihood degree, and establishes the component matching relationship between the target graph model and the candidate graph. On this basis, the (confidence map), of the target location can be obtained by sampling the optimal position of the target. (3) finally, in order to avoid considering only the contribution of the local target components, the discriminant ability is not enough. In order to improve the tracking robustness, we introduce the feature representation of the whole target to participate in the target location voting.
【学位授予单位】：合肥工业大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】