空对地复杂背景目标跟踪算法研究 ——基于SVM目标跟踪算法研究

	空对地复杂背景目标跟踪算法研究 ——基于SVM目标跟踪算法研究
其他题名	Air To Ground Object Tracking in Complicated Background --Study about Object Tracking based on SVM
	郭星辰
学位类型	硕士
导师	张葆
	2015-10
学位授予单位	中国科学院大学
学位专业	光学工程
关键词	支持向量机核函数哈尔特征机器学习
摘要	目标跟踪技术隶属计算机视觉，是图像处理领域研究的热点问题之一，其在战场侦察、电视制导、智能监督等领域有重要应用。传统跟踪算法中，由于目标旋转、尺度变换、遮挡、光照变化以及机载相机对焦、晃动导致的观测信息变化，致使跟踪能力下降。虽然近年来众多学者对目标跟踪技术进行研究，仍无法很好解决上述问题，因此仍具有进一步研究的价值。论文以电视跟踪系统为背景，通过实验分析传统跟踪算法（质心跟踪、相关跟踪）的弊端，提出以支持向量机为框架的分类跟踪算法，重点研究支持向量分类器对目标学习以及重新锁定目标的能力。基于支持向量机对小样本数据学习的能力，将目标跟踪归结于对跟踪波门中特征分类问题。增加跟踪中目标特征类型和数量，提高目标判断唯一性，克服对比度较差的复杂背景环境的问题；通过实时学习更新目标的新特征，兼容相关跟踪的模板刷新问题。基于机器学习理论，将支持向量机设计为闭环学习算法。支持向量机经由手动或检测算法确定的目标训练后，从视频后续建立的测试数据中计算新支持向量添加入原有的训练集中，实现对选定目标的持续学习，从而克服目标旋转、尺度变换等问题。支持向量机自身算法复杂度高，为避免数据冗余造成时间浪费，算法充分利用支持向量机处理高维数据的优势，提取目标及背景的线、边缘、对角、中心等多个特征构成多维行向量，作为支持向量分类器的训练数据。针对空对地目标跟踪中树木、薄雾遮挡等自然物遮挡问题，在行向量中加入全局哈尔特征，利用哈尔特征中减法运算屏蔽均匀遮挡物，实现遮挡物干扰下的准确跟踪。核函数是支持向量机的精髓，在本论文中的应用主要有两个方面：一是利用 I 多种核函数实现输入空间至特征空间的转化，提高支持向量机分类精度，采用IRIS 数据进行同维度向量仿真，将抽象分类过程具体化并确定核函数参数；二是采用核函数对视频序列中图像评价，实现支持向量数据库与图像的快速匹配。经过实验验证，采用仿真确定的核函数参数，在确保高精度跟踪的前提下，能有效的提高的算法的实时性。本论文试图从传统跟踪算法着手，解决跟踪过程中由于目标融合、遮挡等问题导致跟踪精度下降或跟踪失败问题。对分类跟踪算法这一发展趋势进行了研究、分析与展望。
其他摘要	Object tracking belongs to computer vision. As one of the hotspot in the image processing, it plays a great important role in conventional warzone surveillance, TV guidance and intelligent monitor. Precision usually decreases during tracking in traditional tracking algorithm due to the rotation, scale transform and occlusions of target. Furthermore, tracking also can fail because of other change of observation information caused by illumination change, focusing and shake of camera. Above mentioned problems still exists though many scholars make a lot of effort on object tracking. Therefore, object tracking deserves to in-depth study. TV tracker is taken as the research background. This paper presents a framework for tracking by classification based on SVM(Support Vector Machine) with the analysis the disadvantages of traditional tracking algorithm such as geometric centroid tracking and correlation tracking. The ability of learning and relocking target is emphasized in this paper. Tracking algorithm in the paper comes down to features classification. The features are extracted in the tracking gate because the SVM has advantage on handling small samples data. The object in the new images can be confirmed by increasing the number of features in the tracking gate. In this way, object can be tracked accurately in the complicated background. And the new features can be got via online learning. It’s the same with model update in the correlation tracking. SVM is designed as a closed-loop algorithm based on computational learning theory. Training set used for training SVM is made up of object features. This target is designated through manual way or detecting algorithm in the initial video. The testing set is also features, but the object is from the last video. SVM will pick up new SV(Support Vector) from the testing set. And then, these SVs will be added into the training set for another training. In this way , SVM can study the target during the whole video and overcome the problem of the rotation and scale transform. SVM is good for multi-dimensional data, although it costs much time in calculating. Many features of object and background are extracted to form a multi-dimensional data as a row vector, including the line, edge, opposite angles and center. These row vectors comprise the training data. According to the influence of trees and fog in the air to ground tracking, the holistic Haar feature is added into the row vector. Target can be tracked accurately in the case of occlusion, and this is benefited by subtraction of Haar features. Kernel function plays an important role in SVM. It’s mainly used for two ways. One is to transform the input space into feature space whose dimension is higher than the other. The precision of classification is increased in this way and the parameter of kernel function can be confirm at the same time. The abstract classification become materialization via simulation with the help of IRIS. The other way is to evaluate the image in the video. Fast matching between image and data base come true on the basis of assessment result. The experiment proves that the real-time of algorithm is improved adopting the parameter confirmed by simulation, and the object can also be tracker accurately. This algorithm solve the rotation, scale transform and occlusions problems which are found in traditional tracking algorithm. The development tendency of object tracking is study at the end of the paper.
语种	中文
文献类型	学位论文
条目标识符	http://ir.ciomp.ac.cn/handle/181722/49329
专题	中科院长春光机所知识产出
推荐引用方式 GB/T 7714	郭星辰. 空对地复杂背景目标跟踪算法研究 ——基于SVM目标跟踪算法研究[D]. 中国科学院大学,2015.

条目包含的文件
文件名称/大小	文献类型	版本类型	开放类型	使用许可
郭星辰.pdf（2246KB）	学位论文		开放获取	CC BY-NC-SA	请求全文