基于绝对差累加和相关匹配的抗遮挡跟踪算法研究

CORC > 沈阳自动化研究所 > 中国科学院沈阳自动化研究所 > 光电信息技术研究室

题名	基于绝对差累加和相关匹配的抗遮挡跟踪算法研究
作者	冯俊惠
学位类别	硕士
答辩日期	2015-05-26
授予单位	中国科学院沈阳自动化研究所
授予地点	中国科学院沈阳自动化研究所
导师	肖阳辉
关键词	模板提取遮挡判断多子区域相关跟踪目标轨迹预测
其他题名	Research on Anti-occlusion Tracking Arithmetic Based on Sum of Absolute Difference Correlation Matching
学位专业	控制理论与控制工程
中文摘要	目标跟踪是计算机视觉领域中非常重要的研究课题之一，它是指在一段序列图像的每一帧图像中实时地确定目标的位置、姿态、尺寸等几何状态信息，在智能交通、视觉导航、安防监控、精确制导等很多领域都有非常广泛的应用。在精确制导领域，以目前广泛应用的第二代反坦克导弹为例，它主要采用可见光/红外成像复合制导方式，对于成像指令制导技术，需要通过图像处理的方法确定目标位置，再根据目标与瞄准线的偏差角度控制导弹沿瞄准线行驶，直至击中目标。因此，基于图像的目标跟踪算法具有重要的研究意义。但是，导弹发射后，其尾部红外辐射器产生的辐射光斑会遮挡目标，经典的相关跟踪算法会失效。因此，对于抗遮挡跟踪算法的研究在理论方面和实际工程应用方面都有着非常重要的意义。相关跟踪算法能在低信噪比、特征信息较少等条件下对目标进行准确跟踪，绝对差累加和相关匹配算法（即SAD相关跟踪算法）更是具有计算简单、运算量小的优点，但在遮挡发生的情况下该算法会失效。本文在SAD相关跟踪算法的基础上对抗遮挡跟踪算法进行研究，在无遮挡、局部遮挡以及严重遮挡这三种不同情况下采用不同的跟踪策略，实现了基于SAD相关跟踪算法的抗遮挡跟踪。主要研究内容包括： 1) 针对模板提取过程中可能遇到的图像被噪声干扰、目标边界比较模糊或目标被部分遮挡等问题，建立了基于形状保持的CV变分水平集矩形目标分割模型，加入矩形先验形状分割得到包含目标的矩形图像区域，实现了模板自动提取。 2) 当目标未受到遮挡时，采用SAD相关跟踪算法，当目标受到遮挡时，该方法失效，需要改变跟踪策略。为此设计了一种基于SAD的相关跟踪有效性判断方法，根据序列图像SAD矩阵最小值所在的邻域矩阵的相关值，判断遮挡是否发生，提高了跟踪的准确性。 3) 当目标受到局部遮挡时，采用多子区域相关跟踪算法进行目标跟踪。首先在模板划分的基础上确定了匹配模式和遮挡模式，并给出不同匹配模式的SAD计算方法，然后提出了基于距离离散程度和相似度离散程度的匹配模式决策方法以及基于分层最佳匹配模式的位置信息融合方法，实现了局部遮挡情况下的目标跟踪。 4) 当目标受到严重遮挡时，采用基于核岭回归的目标轨迹预测方法进行目标位置预测。该方法通过映射函数将非线性输入映射为高维特征空间中的线性函数，然后根据岭回归原理解算出线性函数的参数，实现了目标位置的预测。为验证本文算法的有效性，在MATLAB平台上进行仿真。实验结果表明，本文算法能够满足跟踪精度和实时性的要求。
索取号	TP391.41/F62/2015
英文摘要	Target tracking is one of key research subjects in computer vision field, it determines real-time target’s location, attitude, dimension and other geometric status information of each frame image in image sequence. Target tracking is widely used in many fields such as intelligent transportation, visual navigation, security surveillance, precision guidance and so on. In the field of precision guidance, take the widely-used second generation anti-tank missile (ATM) for example, it mainly uses visible light/ infrared imaging combined guidance system. Imaging command guidance technology determines target’s position by image processing method, then controls missile traveling along the line of sight according to the deviation angle of target and the line of sight until hitting target. Therefore, target tracking algorithm based on image has important significance. However, after the missile launch, infrared radiator in its tail will produce radiation spot which occludes target, and classical correlation tracking algorithm will fail. Therefore, anti-occlusion tracking arithmetic has important implication in the aspect of theory and actual engineering. Correlation tracking method is accurate when image has low signal-to-noise ratio or target has few feature, in addition, sum of absolute difference correlation matching method (SAD correlation tracking method) has simple and little calculation, however it will be invalid under occlusion. This paper researches on anti-occlusion tracking arithmetic based on SAD correlation tracking method, uses different tracking strategies in three different situations: target isn’t occluded, target is partial occluded, and target is severely occluded, and realizes anti-occlusion tracking. Main contents are as follows: 1) According to the possible problems in the process of extracting template such that target is disturbed by noise, target’s contour is a little vague, target is occluded partially and so on, build the model of rectangle object segmentation based on shape preserving and CV variational level set, obtain a rectangle area including the object by segmentation, and extract the template automatically. 2) When there is no occlusion, use SAD correlation tracking algorithm, and when there is occlusion, the algorithm fails and tracking strategy needs to change. So design a correlation tracking validity judging method based on SAD, select the neighborhood region of minimum value in the SAD matrix of sequential images, calculate correlation value to judge occlusion, so improve accuracy of correlation tracking method. 3) When occlusion occurs, use multi-sub-region correlation tracking arithmetic to track target. First determine matching mode and occluding mode based on template dividing mode, and provide SAD calculation methods of different matching modes; then propose matching mode deciding method based on distance discrete degree and similarity discrete degree; finally propose position information fusing method based on hierarchical best matching mode to fuse position information, so realize target tracking under partial occlusion. 4) When there is severe occlusion, use trajectory predicting method based on kernel ridge regression to predict target’s position. This method maps the nonlinear input to high-dimensional feature space as a linear function through mapping function, then solves parameters of the linear function using ridge regression method, so realizes target’s position prediction. To verify the efficiency of the algorithm proposed in this paper, complete simulation in the MATLAB platform. Experimental results show that, the algorithm can meet the tracking accuracy and real-time requirements.
语种	中文
产权排序	1
页码	67页
内容类型	学位论文
源URL	[http://ir.sia.ac.cn/handle/173321/16773]
专题	沈阳自动化研究所_光电信息技术研究室
推荐引用方式 GB/T 7714	冯俊惠. 基于绝对差累加和相关匹配的抗遮挡跟踪算法研究[D]. 中国科学院沈阳自动化研究所. 中国科学院沈阳自动化研究所. 2015.