复杂地面背景下目标鲁棒跟踪技术研究
发布时间：2018-07-24 08:08
【摘要】：复杂地面背景下目标鲁棒跟踪技术是实现无人机对地侦察打击任务的基础,也是各类精确制导武器在末制导阶段实现实时准确捕获目标的关键。然而由于地面场景中存在光照变化、遮挡、尺度缩放、旋转、非刚体形变、杂乱背景干扰、由物体突然运动引起的图像模糊、相似物干扰等挑战因素,对运动单目标进行鲁棒跟踪仍面临诸多困难。本文从一个典型跟踪系统包含的四个组成部分入手展开研究,以半监督在线学习理论为基础,提出相应的跟踪算法以克服上述因素的影响。主要研究内容包括以下几方面:1、介绍了半监督在线学习的基本理论及两类典型的基于半监督在线学习的目标跟踪方法:Online MIL跟踪与TLD跟踪。它们对开展复杂地面背景下目标鲁棒跟踪技术研究具有重要的指导及借鉴作用。Online MIL跟踪用包袋来封装具有相近标签的实例,并用包袋标签代替实例标签,通过对样本标签进行模糊化处理以弱化监督效应,能较好的解决样本模糊问题,其实质是对目标外观表示方法的创新。TLD跟踪提出了一种新的目标跟踪框架,利用在线学习机制融合跟踪结果与检测结果,同时在线学习也使跟踪算法具有”记忆”功能,当丢失的目标在视场中重现时,能够重新捕获到该目标,它拓展和完善了传统的基于检测的目标跟踪理论,弥补了仅依靠纯检测或纯跟踪方法存在的跟踪性能不稳定的不足,实质是对跟踪方法的创新。2、提出了一种具有尺度自适应的特征压缩跟踪方法,解决辨别式跟踪算法中的样本特征降维及对目标尺度变化的适应性问题。将压缩感知理论引入目标跟踪领域,首先用一个满足有限等距准则的高斯随机测量矩阵对提取的样本特征进行压缩降维,再用降维后的特征进行分类,这不仅有助于降低计算量、提高跟踪算法的实时性,而且由于压缩特征保留了原始特征的大部分信息,因而能较好的表征目标特性、保证目标跟踪精度。同时,为了使跟踪算法适应目标尺度变化,在样本采样阶段,通过结构约束性采样获得能尽量反映目标位置及尺度变化的样本集,以便在跟踪过程中能找到与目标当前状态匹配的最佳样本。3、提出一种基于正负样本响应差异最大化的在线加权特征选择目标跟踪方法,解决特征冗余问题。分类器输入特征数量与输出性能之间不一定存在线性关系,当特征数量超过一定值时,不仅耗费巨大的计算开销,而且还会降低分类器输出性能。通过定义一个样本响应差异函数来选择多个特征选择器(即弱分类器)组成强分类器,并用该强分类器对样本进行分类,分类得分最高的样本块即对应当前帧跟踪结果。在选取特征选择器组成强分类器过程中,根据样本与目标位置间的距离及重叠度关系赋予该样本相应的权重,以突出正样本、抑制负样本,增强分类器对正负样本响应的辨识能力,从而找出最佳正样本来描述目标当前状态。4、提出一种基于压缩特征稀疏表示的目标跟踪方法,解决压缩特征在PCA子空间中的表示问题。PCA子空间表示法用目标模板集的主成分分量来描述候选目标,极大地增强目标外观描述能力,能够克服噪声、光照变化的影响。利用生成式表示策略及增量学习更新方式对表示目标外观模型的压缩特征子空间及琐碎模板进行稀疏表示,将目标跟踪看成是压缩特征的稀疏近似问题。为了更新遮挡条件下的目标外观模型,提出一种逆指示策略,根据压缩特征子空间获得的似然值寻找原始图像空间中具有最大观测似然的图像块。与基于模板集或基于PCA子空间的目标外观表示方法相比,本文方法虽然也需解决一序列l_1正则最小二乘问题,但由于压缩特征维数低,故计算复杂度大大降低。本文方法对光照变化、部分遮挡、尺度及姿态变化等因素的影响具有较强的鲁棒性。5、基于上下文在解决目标抗遮挡、相似表观干扰等方面表现出的优越性,提出两种基于上下文辅助的目标跟踪方法。一种是基于两级隐式形状模型的目标抗遮挡跟踪方法,主要解决严重遮挡条件下的目标定位问题。在两级码本特征中,一级特征源于目标自身,另一级特征源自周围目标,用这些码本特征构建两级投票模型。根据遮挡程度的不同,赋予这些特征不同的投票权重,以提高遮挡条件下的目标定位精度,该方法实质是利用稀疏上下文辅助目标跟踪。另一种是具有尺度及方向自适应的时空上下文(SOASTC)辅助目标跟踪方法,将目标跟踪看成是一个贝叶斯框架下求解目标位置似然置信图极值的问题,其突出的优点在于更新时空上下文模型、获取目标位置似然估计时利用FFT加速运算,运行速度较快。利用主成分分析法求解目标区域权值图像的协方差矩阵,估计目标尺度及旋转角度,自适应目标尺度及方向变化。该方法具有较强的抗遮挡及抗光照变化能力,对目标快速运动引起的图像模糊具有一定的适应能力,能够抗相似物干扰、目标非刚体形变及杂乱背景的干扰,实质是利用稠密上下文辅助目标跟踪。
[Abstract]:The robust tracking technique of target in complex ground background is the basis for realizing the mission of unmanned aerial vehicle to ground reconnaissance. It is also the key to achieve real-time and accurate target acquisition of all kinds of precision guided weapons at the terminal guidance stage. However, because of the illumination changes, occlusion, scale contraction, rotation, non rigid body shape, random background interference, and objects in the ground scene, the object is the key to the realization of real time and accurate target acquisition of all kinds of precision guided weapons. It is still facing many difficulties for the robust tracking of the single target. This paper starts with the four components of a typical tracking system. Based on the semi supervised online learning theory, this paper proposes a corresponding tracking algorithm to overcome the factors mentioned above. The main research contents include the following aspects: 1, the basic theory of semi supervised online learning and two typical target tracking methods based on semi supervised online learning are introduced: Online MIL tracking and TLD tracking. They have important guidance and reference to the research of robust tracking technology for target in complex ground background and.Online MI L tracking uses bags to encapsulate examples with similar labels, and use bag labels instead of instance labels. By fuzzing the sample labels to weaken the supervision effect, it can better solve the problem of sample fuzzy. The essence is to put forward a new target tracking framework for the innovative.TLD tracking of the object appearance representation method, which is used in the application of a new target tracking framework. The line learning mechanism combines the tracking results and the detection results, while the online learning also makes the tracking algorithm have the "memory" function. When the lost target is reproduced in the field of view, it can be recaptured to the target. It extends and perfects the traditional detection based target tracking theory, and makes up for the existence of only pure detection or pure tracking methods. The inadequacy of tracking performance instability is essentially an innovation of the tracking method.2. A feature compression tracking method with scale adaptive is proposed to solve the dimension reduction of sample characteristics and the adaptability to the change of target scale in the discriminant tracking algorithm. The compression perception theory is introduced into the target tracking field, and the first one satisfies the finite element. The Gauss random measurement matrix of the isometric criterion is used to compress and reduce the feature of the extracted sample, then classify the feature after reducing the dimension, which not only helps to reduce the computation and improve the real-time performance of the tracking algorithm, but also preserves most of the information of the original feature because of the compression characteristics, so it can better characterize the target and guarantee the target. At the same time, in order to make the tracking algorithm adapt to the change of target scale, a sample set which can reflect the position and scale of the target as much as possible is obtained by structural constraint sampling in sample sampling stage, so as to find the best sample.3 which matches the current state of the target in the tracking process, and proposes a kind of the most discrepancy based on the positive and negative sample response. The on-line weighted feature selection of the maximization selects the target tracking method to solve the problem of feature redundancy. The number of classifier input features does not have a linear relationship with the output performance. When the number of features exceeds a certain value, it can not only consume huge computation overhead, but also reduce the performance of the classifier. By defining a sample response difference A strong classifier is composed of multiple feature selectors (weak classifier), and the strong classifier is used to classify the samples. The sample block with the highest score is the result of the following frame tracking. In the process of selecting the strong classifier with the selection of feature selectors, the sample is given to the sample according to the relationship between the distance and the overlap between the target location and the target location. In order to highlight positive samples, suppress negative samples and enhance the recognition ability of positive and negative samples, the best positive sample is found to describe the current state of the target.4, and a target tracking method based on compressed feature sparse representation is proposed to solve the representation problem of compression characteristic in the PCA subspace,.PCA subspace representation. The method uses the principal component component of the set of target template to describe the candidate target, greatly enhances the appearance description ability of the target, and can overcome the influence of noise and illumination change. Using the generation representation strategy and incremental learning update method, the compressed feature subspace and trivial template representing the object appearance model are sparse representation, and the target is tracked. In order to update the object appearance model under the occlusion condition, an inverse indicator strategy is proposed to find the maximum likelihood image block in the original image space based on the likelihood value obtained by the compressed feature subspace. Compared with the template based or the PCA subspace based object appearance representation method, This method needs to solve a sequence of l_1 regular least squares problem, but because the dimension of the compression feature is low, the computational complexity is greatly reduced. This method has strong robustness.5 for the influence of illumination change, partial occlusion, scale and attitude change. Based on the above below, the target anti occlusion, similar apparent interference and so on are solved. Two kinds of target tracking methods based on context aided are presented. One is a target anti occlusion tracking method based on two level implicit shape model, which mainly solves the problem of target location under severe occlusion. In the two level codebook feature, the first level feature is derived from the target itself and the other is derived from the surrounding target. The two level voting model is constructed with these codebook features. According to the different occlusion degree, these features are given different voting weights to improve the target location precision under the occlusion condition. The method is essentially using the sparse context to assist the target tracking. The other is the space-time context (SOASTC) aided target with the scale and square adaptation. The tracking method regards target tracking as a problem of solving the maximum likelihood of the likelihood confidence map of the target position under a Bayesian framework. Its outstanding advantage is to update the spatio-temporal context model and obtain the target location likelihood estimation by using the FFT acceleration operation. The principal component analysis method is used to solve the co square of the target area weight image. Difference matrix, estimation of target scale and rotation angle, adaptive target scale and direction change. This method has strong ability to resist occlusion and illumination change. It has certain adaptability to image blurring caused by fast moving target, can resist the interference of similar objects, the object is non rigid body shape change and the disturbance of random background, in essence it is the use of consistency. Dense context assisted target tracking.
【学位授予单位】：国防科学技术大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：V249;TJ765;TP391.41
本文编号：2140757
[Abstract]:The robust tracking technique of target in complex ground background is the basis for realizing the mission of unmanned aerial vehicle to ground reconnaissance. It is also the key to achieve real-time and accurate target acquisition of all kinds of precision guided weapons at the terminal guidance stage. However, because of the illumination changes, occlusion, scale contraction, rotation, non rigid body shape, random background interference, and objects in the ground scene, the object is the key to the realization of real time and accurate target acquisition of all kinds of precision guided weapons. It is still facing many difficulties for the robust tracking of the single target. This paper starts with the four components of a typical tracking system. Based on the semi supervised online learning theory, this paper proposes a corresponding tracking algorithm to overcome the factors mentioned above. The main research contents include the following aspects: 1, the basic theory of semi supervised online learning and two typical target tracking methods based on semi supervised online learning are introduced: Online MIL tracking and TLD tracking. They have important guidance and reference to the research of robust tracking technology for target in complex ground background and.Online MI L tracking uses bags to encapsulate examples with similar labels, and use bag labels instead of instance labels. By fuzzing the sample labels to weaken the supervision effect, it can better solve the problem of sample fuzzy. The essence is to put forward a new target tracking framework for the innovative.TLD tracking of the object appearance representation method, which is used in the application of a new target tracking framework. The line learning mechanism combines the tracking results and the detection results, while the online learning also makes the tracking algorithm have the "memory" function. When the lost target is reproduced in the field of view, it can be recaptured to the target. It extends and perfects the traditional detection based target tracking theory, and makes up for the existence of only pure detection or pure tracking methods. The inadequacy of tracking performance instability is essentially an innovation of the tracking method.2. A feature compression tracking method with scale adaptive is proposed to solve the dimension reduction of sample characteristics and the adaptability to the change of target scale in the discriminant tracking algorithm. The compression perception theory is introduced into the target tracking field, and the first one satisfies the finite element. The Gauss random measurement matrix of the isometric criterion is used to compress and reduce the feature of the extracted sample, then classify the feature after reducing the dimension, which not only helps to reduce the computation and improve the real-time performance of the tracking algorithm, but also preserves most of the information of the original feature because of the compression characteristics, so it can better characterize the target and guarantee the target. At the same time, in order to make the tracking algorithm adapt to the change of target scale, a sample set which can reflect the position and scale of the target as much as possible is obtained by structural constraint sampling in sample sampling stage, so as to find the best sample.3 which matches the current state of the target in the tracking process, and proposes a kind of the most discrepancy based on the positive and negative sample response. The on-line weighted feature selection of the maximization selects the target tracking method to solve the problem of feature redundancy. The number of classifier input features does not have a linear relationship with the output performance. When the number of features exceeds a certain value, it can not only consume huge computation overhead, but also reduce the performance of the classifier. By defining a sample response difference A strong classifier is composed of multiple feature selectors (weak classifier), and the strong classifier is used to classify the samples. The sample block with the highest score is the result of the following frame tracking. In the process of selecting the strong classifier with the selection of feature selectors, the sample is given to the sample according to the relationship between the distance and the overlap between the target location and the target location. In order to highlight positive samples, suppress negative samples and enhance the recognition ability of positive and negative samples, the best positive sample is found to describe the current state of the target.4, and a target tracking method based on compressed feature sparse representation is proposed to solve the representation problem of compression characteristic in the PCA subspace,.PCA subspace representation. The method uses the principal component component of the set of target template to describe the candidate target, greatly enhances the appearance description ability of the target, and can overcome the influence of noise and illumination change. Using the generation representation strategy and incremental learning update method, the compressed feature subspace and trivial template representing the object appearance model are sparse representation, and the target is tracked. In order to update the object appearance model under the occlusion condition, an inverse indicator strategy is proposed to find the maximum likelihood image block in the original image space based on the likelihood value obtained by the compressed feature subspace. Compared with the template based or the PCA subspace based object appearance representation method, This method needs to solve a sequence of l_1 regular least squares problem, but because the dimension of the compression feature is low, the computational complexity is greatly reduced. This method has strong robustness.5 for the influence of illumination change, partial occlusion, scale and attitude change. Based on the above below, the target anti occlusion, similar apparent interference and so on are solved. Two kinds of target tracking methods based on context aided are presented. One is a target anti occlusion tracking method based on two level implicit shape model, which mainly solves the problem of target location under severe occlusion. In the two level codebook feature, the first level feature is derived from the target itself and the other is derived from the surrounding target. The two level voting model is constructed with these codebook features. According to the different occlusion degree, these features are given different voting weights to improve the target location precision under the occlusion condition. The method is essentially using the sparse context to assist the target tracking. The other is the space-time context (SOASTC) aided target with the scale and square adaptation. The tracking method regards target tracking as a problem of solving the maximum likelihood of the likelihood confidence map of the target position under a Bayesian framework. Its outstanding advantage is to update the spatio-temporal context model and obtain the target location likelihood estimation by using the FFT acceleration operation. The principal component analysis method is used to solve the co square of the target area weight image. Difference matrix, estimation of target scale and rotation angle, adaptive target scale and direction change. This method has strong ability to resist occlusion and illumination change. It has certain adaptability to image blurring caused by fast moving target, can resist the interference of similar objects, the object is non rigid body shape change and the disturbance of random background, in essence it is the use of consistency. Dense context assisted target tracking.
【学位授予单位】：国防科学技术大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：V249;TJ765;TP391.41
【参考文献】
相关期刊论文 前2条
1 石光明;刘丹华;高大化;刘哲;林杰;王良君;;压缩感知理论及其研究进展[J];电子学报;2009年05期
2 董乐红;耿国华;高原;;Boosting算法综述[J];计算机应用与软件;2006年08期
相关硕士学位论文 前1条
1 屈冉;压缩感知算法及其应用研究[D];南京邮电大学;2013年
，本文编号：2140757
本文链接：https://www.wllwen.com/kejilunwen/hangkongsky/2140757.html
