【作者】 李莉；

【导师】 薛雅丽；

【作者基本信息】 南京航空航天大学 ， 控制理论与控制工程， 2014， 硕士

【摘要】 随着红外成像技术的不断发展以及现代战争的需要，红外图像中弱小目标的检测与跟踪技术已成为当今世界红外精确制导研究中的主流方向。弱小目标的检测与跟踪性能的好坏将直接决定制导系统的有效作用距离和设备的复杂程度，它在红外成像制导系统中发挥着举足轻重的作用。本文对红外弱小目标检测与跟踪中图像预处理、目标检测和目标跟踪三个主要技术环节进行了系统的研究。首先，从时域、空域、频域滤波等方面分析了目前较为常用的图像预处理方法，在此基础上引入视觉注意机制，提出了一种全新的基于视觉显著性特点的图像预处理算法。算法通过对图像中对象的灰度显著性分析，快速搜索整个输入图像以得到感兴趣的候选目标集合。实验表明，新算法能够有效抑制红外图像背景，且与大部分空域算法不同，算法不需要进行背景预测，不依赖于目标的先验知识，因而适用范围更广。其次，针对传统单帧检测算法检测概率低、算法鲁棒性差、实时性差的问题，提出了一种基于多方向复合窗结构的单帧检测算法。该算法对红外图像进行灰度特性分析，通过构建窗结构对窗内部像素进行多方向灰度均值的比较来检测目标，不需进行二值分割，避免了检测的结果过分依赖于阈值的选取。最后，为了克服传统Mean-shift算法进行红外弱小目标跟踪时存在目标描述不稳健，跟踪精度不高的问题，提出了联合多特征的Mean-shift红外弱小目标跟踪方法。该方法采用分层跟踪的思路，首先对Mean-shift算法进行改进，然后将改进后的Mean-shift算法与Harris特征匹配方法相结合，在Mean-shift算法得到当前帧红外图像中目标的位置后，采用Harris特征匹配方法对该位置坐标进行修正，从而保证了目标跟踪的精确性。其中改进的Mean-shift算法采用多个目标特征来建立目标模板，并且为了使跟踪鲁棒性更强，考虑了目标模板实时更新的问题。实验表明，稳健的目标描述加上特征匹配对跟踪结果的准确修正，能够有效消除跟踪时产生的偏离误差，从而提高了跟踪的精度，确保了目标的准确跟踪。更多还原

【Abstract】 With the development of infrared imaging technology and the need of modern warfare, detectionand tracking technology of dim and small target in infrared image has become mainstream directionof infrared precision guidance technology in modern world. The quality of detection and trackingperformance of dim and small target directly determines the effective distance of guidance system andthe complexity of the device, and it plays a pivotal role in the infrared imaging guidance system. Inthis paper, three main technical aspects of dim and small target detection and tracking including imagepre-processing, object detection and object tracking are studied.First of all, the commonly used image pre-processing methods are analyzed from the timedomain filtering, spatial domain filtering, and frequency domain filtering in this paper. Visualattention mechanism is introduced on this basis, and a new image preprocessing algorithm based onvisual significance characteristics is also proposed in our paper. The algorithm rapidly search theentire input image to obtain a set of interested candidate targets by analysis the grayscale significanceof the objects in the image. Experimental results show that the new algorithm can effectively inhibitthe infrared image background. The new algorithm which is different from the majority of spatialdomain algorithm does not require background predict, and does not depend on prior knowledge,therefore it applies more broadly.Secondly, on account of low probability of detection, poor robustness, poor real-time problemsof the traditional single-frame detection algorithm, a detection algorithm based on multi-directionalcomposite window structure combined with the preprocessing framework of the last chapter isproposed. The gray level distribution of infrared image is analyzed in this algorithm. To detect thetarget by building window structure to compare pixels inside the windows through multi-directionalgrayscale mean value. Without binary segmentation, this algorithm avoids the detection resultsexcessively depend on the threshold selecting.Finally，to overcome the disadvantages of traditional Mean-shift algorithm such as the instabilityof target description and the low accuracy of tracing, an improved Mean-shift algorithm based onfusing multiple features is proposed. The hierarchical tracking method is used in this algorithm. Firstimproving the Mean-shift algorithm, and then combining the Harris feature matching algorithm withthe improved method. After the Mean-shift algorithm obtaining the position of the target in the currentinfrared images, we use the Harris feature matching method to correct the position coordinates, thus ensuring the tracking accuracy. The improved Mean-shift algorithm adopts several targetcharacteristics during target description, and considers the real-time updating problem of targettemplate in order to make the tracking strong robustness. Experimental results show that robust objectdescription, with the accurate correction to the tracking results of the feature matching, can effectivelyeliminate the deviation errors generated in the process of tracking, thus improving the accuracy oftracking and ensuring the accurate tracking of the target.更多还原