在线半监督红外跟踪算法研究

发布时间：2018-05-20 16:12

本文选题：半监督学习 + 在线学习　；参考：《华中科技大学》2016年博士论文

【摘要】：本课题来源于某**红外跟踪平台的预研项目。在线红外跟踪是一个极具挑战性的问题,在精确制导、飞行控制、无人机侦察、安防智能视频监控、基于视频的人机交互以及智能视觉导航中都有着广泛的应用。由于红外图像相对于可见光图像大多对比度较低,因此现有的大量视觉跟踪方法在实际应用时难以对红外图像中目标进行快速准确的检测和鲁棒的跟踪。本文采用了近些年兴起的基于检测的跟踪系统作为基本的红外跟踪模型。针对该跟踪模型中的采样标记环节,专注于研究通过减少标记信息的错误来减弱在红外跟踪过程中出现的漂移问题,从而实现鲁棒的在线红外跟踪。本文的主要贡献如下：首先,本文基于充分性降维和半监督学习首次提出了充分性半监督特征提取理论,并设计了充分性半监督红外特征提取算法(fusion refinement:FR算法和semi-supervised kernel fusion refinement:SemKFR算法)。前一种算法提取的是线性特征,后一种算法提取的是非线性特征。相比于其他特征提取算法,本文提出的这两种算法能够同时利用部分置信度高的标记信息和大量未标记信息,且以最小化样本信息损失为目标来提取样本特征,因此提取的特征具有更强的区别性。文中大量红外图像的实验结果也证实了这两种充分性半监督特征提取算法的特征提取能力。其次,研究了基于半监督粗糙公共向量(SRCV)的在线红外跟踪算法。该算法继承了半监督学习能够部分标记样本的特性和判别公共向量(DCV)算法对于小样本问题的鲁棒性。在红外跟踪的初始阶段,由于样本匮乏,每个样本(图像块)维度很高,这是典型的小样本问题。经研究发现,本文提出的SRCV算法在样本较少时仍能够学习红外目标的低维特征。同时为了适应在线红外跟踪,本文在SRCV算法的基础上提出了增量式的SRCV在线特征提取算法(ISRCV).ISRCV算法利用随机投影树(RPTree)来近似红外跟踪过程中增量式采集样本的内部流形结构,同时借助于RPTree构造了用于在线学习的目标函数,并提出了迭代更新准则求解算法。实验结果及分析表明,ISRCV提取的在线特征能够有效保存红外目标的主要特征,在实验中取得了与现有算法相近的跟踪效果。再次,设计了基于边缘信息的半监督红外特征表达算法(增量半监督推广的公共向量分析算法：ISSGCVA)。在利用基于检测的跟踪系统框架设计在线红外目标跟踪的过程中,即使是采样少量的标记样本也可能会存在标记误差。本文提出的ISSGCVA算法不需要指定每个样本的具体标记信息,而只用给定相似和不相似样本对,就可以进行特征的学习,从而能够进一步减小标记信息对于红外跟踪漂移的影响。同时,ISSGCCVA算法放宽了投影向量严格位于相似离散矩阵的零空间的约束,通过这种方式,本文的ISSGCVA算法既能够处理红外跟踪初期的小样本情形,又能够处理较长时间跟踪后的大样本情形。此外,本文还详细推导了一种有效的迭代算法来快速求解ISSGCVA算法的目标函数。最后本文基于边缘信息的ISSGCVA算法提出了一个在线红外跟踪系统,并在实验中证实了它的有效性。接着,本文针对基于检测的跟踪系统框架中的分类器模块进行了研究,探索了一种改进的半监督增量可变流形嵌入(ISemFME)分类算法。ISemFME算法继承了半监督学习能够同时利用标记和未标记样本进行分类模型的学习,而且能够在线的更新分类算法的参数,以适应红外跟踪过程中目标外观的不断变化。同时,考虑到目标外形的多样性,ISemFME算法引入了回归误差函数来构造目标函数,并且证明了该目标函数是凸函数,可以解析求解。为了适应在线红外跟踪,本文还提出了缓冲策略(buffering strategy)来降低ISemFME算法的时间复杂度和空间复杂度。在VOT-TIR2015红外数据库上,基于ISemFME算法跟踪系统表现出了较低的时间复杂度和很高的跟踪准确率。最后,由于红外小目标缺乏一定的外观形状,很多的特征提取算法并不能对红外小目标有很好的效果。本文针对红外小目标提取改进了一种融合图像增强和目标提取的算法—增强一比特变换(Enhanced one-bit transform:En1BT)算法,并且在实验中验证了该算法的有效性。本文围绕基于检测的跟踪系统,研究了红外跟踪过程中的两个关键问题：如何提取鲁棒的红外目标特征以及怎样在观测噪声下构造分类函数,提出了基于半监督学习的在线红外跟踪系统,并在大量仿真实验中得到了证明。本文所提出的在线半监督红外跟踪理论、模型和算法对于其他的计算机视觉理论及应用也有指导意义。
[Abstract]:This topic comes from a pre research project of an infrared tracking platform. The online infrared tracking is a very challenging problem. It has extensive applications in precision guidance, flight control, UAV reconnaissance, security intelligent video surveillance, video based human-computer interaction and intelligent visual navigation. Most of the existing visual tracking methods are difficult to fast and accurate detection and robust tracking of the target in the infrared image. In this paper, a tracking system based on detection in recent years is used as the basic infrared tracking model. The main contributions of this paper are as follows: firstly, this paper first proposed a sufficient semi supervised feature extraction theory based on sufficient reduced dimension and semi supervised learning, and designed the adequacy. The semi supervised infrared feature extraction algorithm (fusion refinement:FR algorithm and semi-supervised kernel fusion refinement:SemKFR algorithm). The previous algorithm extracts linear features and the latter algorithm extracts nonlinear characteristics. Compared with other feature extraction algorithms, the two algorithms proposed in this paper can use partial confidence simultaneously. High markup information and a large number of unlabeled information, and to minimize sample information loss as the target to extract the sample features, so the features extracted are more distinct. The experimental results of a large number of infrared images in this paper also confirm the feature extraction capability of the two full semi supervised feature extraction algorithms. Secondly, the study is based on semi supervision. The online infrared tracking algorithm of the governor rough common vector (SRCV). This algorithm inherits the characteristics of semi supervised learning which can partially mark the sample and discriminate the robustness of the common vector (DCV) algorithm for small sample problems. In the initial phase of the infrared tracking, the dimension of each sample (image block) is very high because of the lack of samples. This is a typical small sample. It is found that the proposed SRCV algorithm can still learn the low dimensional features of infrared targets when the sample is small. In order to adapt to the online infrared tracking, an incremental SRCV online feature extraction algorithm (ISRCV).ISRCV algorithm is proposed on the basis of the SRCV algorithm, and the random projection tree (RPTree) is used to approximate the infrared tracking. In the process, the internal manifold structure of the sample is incrementally collected. At the same time, the target function for online learning is constructed with the help of RPTree, and an iterative updating criterion is proposed. The experimental results and analysis show that the online features extracted by ISRCV can effectively preserve the main characteristics of the infrared target, which is similar to the existing algorithms in the experiment. Thirdly, a semi supervised infrared feature expression algorithm based on edge information (incremental semi supervised generalized vector analysis algorithm: ISSGCVA) is designed. In the process of designing an online infrared target tracking based on a detection based tracking system framework, even a small number of labeled samples may also have markup errors. The ISSGCVA algorithm proposed in this paper does not need to specify the specific labeling information of each sample, but only a given similar and dissimilar sample pair can be used for characteristic learning, which can further reduce the influence of the label information on the infrared tracking drift. At the same time, the ISSGCCVA algorithm relaxes the projection vector to be strictly located in the similar discrete matrix. In this way, the ISSGCVA algorithm in this paper can not only deal with small sample cases in the initial stage of infrared tracking, but also can handle large sample cases after a long time tracking. In addition, an effective iterative algorithm is also derived to quickly solve the target function of the ISSGCVA algorithm. Finally, this paper is based on the edge information. The ISSGCVA algorithm proposed an online infrared tracking system and proved its effectiveness in the experiment. Then, this paper studied the classifier module in the framework of detection based tracking system, and explored an improved semi supervised incremental variable manifold embedding (ISemFME) algorithm.ISemFME algorithm to inherit the semi supervised learning. At the same time, learning can use markers and unlabeled samples for classification model learning, and can update the parameters of the classification algorithm online to adapt to the constant changes in the appearance of the target in the infrared tracking process. At the same time, considering the diversity of the target shape, the ISemFME algorithm introduces the return error function to construct the target function, and proves that the target function is constructed. The objective function is a convex function, which can be solved analytically. In order to adapt to the online infrared tracking, this paper also proposes a buffer strategy (buffering strategy) to reduce the time complexity and space complexity of the ISemFME algorithm. On the VOT-TIR2015 infrared database, the ISemFME algorithm tracking system shows a lower time complexity and high degree. In the end, because of the lack of a certain shape of the small infrared target, many feature extraction algorithms do not have a good effect on the small infrared targets. In this paper, an improved algorithm of image enhancement and target extraction, enhanced Enhanced one-bit transform:En1BT algorithm, is improved for infrared small target extraction. In this paper, the effectiveness of the algorithm is verified in the experiment. In this paper, two key problems in the infrared tracking process are studied around the detection based tracking system: how to extract the robust infrared target features and how to construct the classification function under the observed noise, and put forward an online infrared tracking system based on semi supervised learning. It is proved that the on-line semi supervised infrared tracking theory, the model and the algorithm are also instructive to other computer vision theories and applications.
【学位授予单位】：华中科技大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TN219;TP391.41

【相似文献】