【作者】 高文；

【导师】 朱明；

【作者基本信息】 中国科学院研究生院（长春光学精密机械与物理研究所） ， 机械电子工程， 2012， 博士

【摘要】 从19世纪50年代最早的机载平台的雏形（用风筝、气球将相机带到天上对地拍摄）开始，到20世纪初用飞机进行照相侦查，到二战期间航空相机作为机载平台的一种并成为一种主要的航空侦查工具，再到20世纪中后叶机载平台实现了图像传输和目标跟踪功能使光电平台进入另一个时代，到现在机载光电平台已经具有了侦查、跟踪、导航、图像实时传输、测距、目标定位等众多功能，成像设备也从单一的光学照相变为多光谱、红外、微波、合成孔径雷达等多种形式，其应用领域更是已经涉及到视频转播、测绘、资源调查、环境监测、巡逻、地质勘探、空间科学试验、航空侦查、军事打击等众多领域。而目标跟踪技术作为机载平台实现目标侦察、监控、跟踪打击的关键技术，也是计算机视觉领域的一个重要方向，它是一项融合图像处理、模式识别、人工智能、自动控制等多种不同领域先进成果的课题。本文是以“机载光电平台目标跟踪技术”为研究内容的，即主要研究针对机载平台的目标跟踪技术，而针对机载平台尚存在着很多难题，诸如平台抖动造成的图像模糊、目标大角度旋转、三维仿射变换、大尺度缩放、平台和目标的相对运动造成的视场变化、目标自身外观的改变、非刚体运动时的形变、部分或全部阻塞以及跟踪丢失后的重新捕获目标等等，都是亟待解决的问题，从而才能实现长期、稳定的跟踪，除了上述鲁棒性问题还有一个非常重要的问题就是实时性这也是能不能在平台上应用的重要指标之一。针对上面所说的问题，很多学者利用一些不变性特征（如SIFT(Scale InvariantFeature Transform)、SURF(Speeded Up Robust Features)等）实现了在不同尺度、旋转等情况下的特征提取与匹配，还有一些研究人员利用均值漂移理论实现了目标旋转、光照变化、部分遮挡等情况的目标跟踪，利用卡尔曼滤波、粒子滤波实现在部分或全部遮挡情况下利用运动模型预测出目标位置也是一种办法，但是这些算法都不能针对上面所述的所有难题予以解决，都是在某些情况下适用，本文针对这些问题展开了研究。首先是研究了对目标的建模，现在可以大体的将目标建模分成两大类，一类是注重特征本身的不变性（即在尺度变化、旋转等情况下能够有很好的重复性），这类方法在特征位置的选择、特征点的表示两大方面，都要花费巨大的计算量，以求得不变性，因其特征点的表示复杂所以在特征点匹配上也同样需要花费较多的计算量和时间；另一种方法是并不注重单个特征点的这些不变性特性，而是通过一些统计理论使这种位置选取和特征表示都相对简单的特征点联合使用，已达到鲁棒的效果，本文第二章主要研究了这两个类型的特征即目标建模。第三章则研究了基于这种局部不变性特征的匹配实现目标跟踪，首先是采用KLT(Kanada Lucas Tomasi)和RANSAC(RANdom Sample Consensus)算法实现了图像的匹配，其可以用来实现稳像；然后则研究了基于SURF特征点匹配的目标跟踪，并针对高维特征点的搜索匹配过程运算量大耗时长的情况采用期望最大(Expectation-Maximization, EM)算法进行匹配过程，并达到了很好的实验效果。第四章开始，本文着重研究了统计学习中的半监督学习理论，阐述了半监督学习与监督学习、非监督学习的区别，以及其利用未标记数据对于学习器性能提升的作用等，最后在学习了众多理论的基础上提出一种步步为营的反馈式学习方法（Bootstrapping Feedback Learning，BFL），并证明了在符合一定条件下利用未标记数据通过提出的约束的反馈式学习能够步步为营的提升学习性能。第五章与第四章有着紧密的联系，首先使用随机蕨类分类器（Fern）与BFL相结合实现了基于Fern-BFL的目标检测，而后提出一种目标跟踪的通用框架，将传统的只有跟踪过程或者只有检测过程或者目标更新机制简单等进行了颠覆，并采用Fern-BFL对其进行了实现，通过3部分的实验表明该算法能够实现长期稳定的跟踪，有非常好的鲁棒性，及目标重新捕获能力。第六章则是针对实时性兼顾一定的鲁棒性进行了研究，提出采用粒子群优化加速Mean-Shift的算法，实现了快速的目标跟踪满足实时性要求，并通过实验证明其在目标旋转、模糊、有相似性目标时也有很好的鲁棒性。更多还原

【Abstract】 Since the19th century50’s, the earliest prototype of aircraft-borne platform(people used kite and balloons to bring camera to the sky) was come out, and at thebeginning of the20thcentury people use airplane for photography investigation, andduring the world war II aviation camera as a form of aircraft-borne platform hadbecome a major air detection tools, then the late of20thcentury, aircraft-borneplatform had already had the function of image transmission, and target tracking,which had entered to another epoch, and today the present aircraft-bornephotoelectricity platforms have multitudinous functions such as investigation, andtrack, and navigation, and image real-time transmission, and Range finder, and theimaging device has changed from single optical camera to many kinds of forms suchas multi-spectral, infrared, microwave, and synthetic aperture radar, what’s more, itsapplication area has involved into the video broadcast and retransmission, surveyingand mapping, resource investigation, environmental monitoring, patrolling, andgeological prospecting, scientific experiments in space, air reconnaissance, militarycombat, and many other fields.And target tracking technology as the key technology for airborne platform torealize those functions (target detecting, monitoring, tracking, and attacking), is anadvanced topic which syncretizes image processing, the pattern recognition, theartificial intelligence, the automatic control and many kinds of different domainadvanced achievement. This paper takes “the target tracking technology on the aircraft-bornephotoelectricity platform” as the research content, namely, main research is targettracking applied at the aircraft-borne photoelectricity platform, in this field, there arestill many challenges, such as image fuzzy caused by platform jitter, large anglerotating of object,3D affine transformation, large-scale zoom, change of the field ofview caused by the relative movement between target and platform, change of goal’sappearance, non-rigid deformation, partially or completely occlusion, and the targetacquisition after tracking failure, those are all the questions which urgently await to besolved, then the long-term stable tracking will come true. However, except the aboverobustness, another important issue is real time which is one of major indicatorwhether an algorithm can be used in platform.For the problem described above, many scholars make use of some immutabilityfeatures (such as SIFT, SURF) to realize feature extractions and matching in differentscales, and rotation; and some other ones research Mean-Shift theory and realize thetarget tracking under the case of target revolving, illumination change, part occlusion;and using the Kalman filtering, the particle filter theory, to realize forecasting targetposition in the situation of part or complete occlusion also is one means, but none ofthese algorithms can address all of the above-mentioned challenges, they are allsuitable in some circumstances, this paper focuses on research on these issues.The first is the study of the objectives of the modeling, presently; it can generallybe divide into two categories, one of which is invariability feature (which has goodrepetitiveness even in the situation of scale change and rotation), method of this kindin the step of extracting feature and depicting the feature cost huge computation, andthe feature matching step also spend more calculation; the other one is simple featurewhich do not pay much attention to the invariance of single feature, use statisticaltheory to make the simple feature which is easy to find and depict joint use, thenachieve the robust result, the second chapter mainly researched this two style featurewhich is target modeling.In chapter3, we did some research on target tracking with local invariancefeature matching. First, we use KLT and RANSAC algorithm realize image matching, which can be used in image stabilization; second, target tracking with SURF, and forhuge computation and long time taken by searching process of high dimension featuredepicter, we try to use expectations maximum (EM) algorithm to feature matchingstep, and achieved the very good experimental results.In Chapter4, this article focuses on research in semi-supervised learning whichis the part of the statistical learning theory, and elaborates the difference betweensemi-supervised learning, supervised learning and non-supervised learning, and theuse of unsigned data improving the learner performance; finally, on the basis ofstudying many theory, we propose a boostrapping feedback learning (BFL) method,and had proofed that once meet several conditions, feedback learn using unsigned datacan improve the performance at every step.Fifth chapter has the close relation with the fourth chapter, first, we userandomize fern and BFL combined realize target detection based on Fern-BFL, thenpropose a framework of general target tracking system, which subverts traditionalonly tracking or detecting and simple target model exchange mechanism. Then we useFern-BFL on its implementation, by3parts experiments we conclude that this targettracking system has good robustness and the capabilities of re-capture, and cantracking long-term and stably.A method using partical swarm optimism (PSO) accelerate Mean-Shift algorithmfor a real-time balance for a robust of tracking was researched in the sixth (6th)chapter, which achieves fast target tracking and satisfies the real-time request, throughexperiments we conclude that it’s robust to object revolving, fuzzy, similar objectnearby.更多还原