【作者】 钟权；

【导师】 吴钦章； 周进；

【作者基本信息】 中国科学院研究生院（光电技术研究所） ， 信号与信息处理， 2014， 博士

【摘要】 扩展目标跟踪技术是图像处理领域的一个重要组成部分，在现实生活中也有较强的应用前景。但是，由于目标及其所处背景的多样性，使得没有一种有效的算法可以解决所有的跟踪问题。同时，目标姿态变化、旋转变化、运动模糊及其光照变化等因素的存在，也给跟踪算法的研制带来了困难。目前，虽然有很多成功经典的算法被陆续提出，但是仍有很多问题没有解决。本文针对这些问题，分析了一般目标跟踪算法的组成模块和处理流程以及扩展目标稳定跟踪技术的难点，并且研究了两种经典跟踪算法的优势和缺点。在此基础上，对扩展目标的稳定跟踪技术进行了深入研究。主要包括三个方面：首先，深入研究特征对扩展目标跟踪算法性能的影响。从特征选择和融合的角度解决扩展目标跟踪问题。文章对扩展目标所具有的、较为常用的特征进行了介绍。对于单个特征不足以对目标进行有效跟踪的问题，考虑使用多个特征融合跟踪。分析了多种特征级别的融合方式。并在此基础上，提出一种基于目标轮廓直线的参数方程和边缘灰度直方图相结合的跟踪算法。首先，利用Hough变换得到待匹配直线的直线参数方程，同时建立该直线边缘的归一化灰度直方图，由此得到的初始模板包含了目标的空间位置和像素分布信息。然后，在后续帧中，利用梯度方向和参数方程对分割后的像素进行聚类，得到多条待匹配直线。最后，利用灰度直方图匹配算法获得直线边缘的精确位置。实验证明该算法对处于复杂背景下、存在局部短时遮挡的、包含直线边缘的目标有很好的跟踪效果。其次，深入研究模板匹配算法中的关键技术点，包括模板的创建和模板的更新。从解决尺度问题的角度来阐述扩展目标跟踪算法的研究。本文从目标的灰度信息分布特点出发，提出一种基于加权信息熵的初始位置修正算法，首先，在搜索窗口中获取测试样本，然后，计算各个样本的加权信息熵，接着，通过先验信息对样本进行筛选，获得熵值最小区域，从而得到修正后的目标位置。从背景和目标的可区分性上来验证算法的有效性，实验结果表明，该算法对于处在复杂背景下的目标能够正确、可靠稳定的对其位置进行修正。对于模板尺度修正问题，使用SIFT特征点在前后两帧中的匹配关系求得模板变化的仿射变换参数，由此，对模板尺度进行自适应更新。在Meanshift算法框架下测试算法效果，实验证明，SIFT特征可以有效解决模板尺度更新问题。最后，深入研究稀疏特征和目标外观模型对扩展目标跟踪算法性能的影响。从目标外观模型的角度阐述扩展目标跟踪问题。压缩跟踪算法作为一种新的算法，具有简单、高效、实时的优点，但该算法依然存在缺陷。首先，在复杂背景或有遮挡等情况下，容易较快的引进误差；其次，跟踪窗口保持不变，使得不能正确跟踪目标位置且不能准确更新正负样本；最后，搜索样本数目大，导致跟踪速度不理想。针对这些问题，利用前后帧跟踪点的直方图对比来判断遮挡的发生，并自适应的改变更新系数；采用在原算法最优匹配点周围小范围多尺度搜索更优位置的方法，来适应目标尺寸的变化；引入粗精跟踪策略，在不同阶段使用不同数量的子特征集进行匹配，以筛选样本、减少计算量。这些改进避免了算法缺陷导致的跟踪失败，提高了跟踪效率。实验证明，改进后的算法比原算法具有更好的鲁棒性且跟踪速度更快。同时，进一步提出一种新的融合尺度不变特征和压缩特征的目标跟踪算法以解决姿态变换、光照变化、旋转和运动模糊下目标的稳定准确跟踪问题。算法使用压缩特征对目标和背景进行描述，通过在图像帧中采集到的正负样本在线训练和学习支持向量机分类器，将跟踪任务构建为一个二类分类问题。使用该分类器对下一帧的目标和背景进行分类，以从而获得精确的目标位置和区域。同时，算法使用前后两帧的尺度不变特征特征点之间的对应匹配关系求解目标尺寸变化值，从而实现模板大小的自适应调整。将算法与其他算法在某些图像序列上的跟踪比较显示，该算法在有效性、正确性和鲁棒性上性能优越。更多还原

【Abstract】 Extended target tracking technology is an important part of the field of imageprocessing. There are strong prospects in real life. However, due to the diversity ofthe target and its background which makes no effective tracking algorithm can solveall problems. Simultaneously, it is a challenging task to develop efective andefcient appearance models for robust object tracking due to factors such as posevariation, illumination change, occlusion, and motion blur. While much success andclassical algrithms has been demonstrated, numerous issues remain to be addressed.Aiming at these problems, analyse the blocks and processes of the general targettracking algorithm, as well as difficulties in stabilizing extended target trackingtechnology, and studied two classical tracking algorithm’ advantages anddisadvantages, and on this basis, the stability of the extended target trackingtechniques were studied. Mainly includes three aspects:Firstly, in-depth study of the impact of the expansion feature target trackingalgorithm performance. Solve the problem of extended target tracking featureselection and fusion perspective. Many commonly features were introdeced. For asingle feature is insufficient to effectively target tracking problem, consider usingmultiple feature fusion tracking. Analyse the way of fusion, and on this basis, anonline combined two kinds of features method for visual object tracking wasproposed. A feature set was built by combining object contour’s linear parametricequation and edge histogram. Firstly, the Hough transform is used to detect line inpicture and obtain the parameters of the line, meanwhile, get the histogram at thesame position of the line. The resulting initial template contains the location andpixel distribution information of the target. Secondly, in the subsequent frames,pixels divided and clustered by gradient direction and parameter equation wouldform more than one line which was to be matched. At last, obtain the exact locationof the straight edge using the histogram matching algorithm. The experimentalresults indicate that the algorithm can run effectively when occlusion andcomplicated backgrounds happened, and perform favorably on challengingsequences in terms of efficiency, accuracy and robustness.Secondly, in-depth study of the template creation and updating. To solve theproblem from the perspective of the scale change. A modified initial position algorithm based on the weighted information entropy was proposed. At first, obtaintest samples in the search window. Then, calculate the weighted information entropyof each sample. Next, the sample is filtered out by priori information to obtain theminimum entropy region and the corrected target position. According to thedifferences beteen background and objectives to distinguish and verify theeffectiveness of the algorithm. Experimental results show that the algorithm is in thetarget complicated background can correct, reliable and stable correction of itsposition.For the problem of fixing template scale, use the SIFT feature matchingbetween two frames to obtained the affine transformation parameters which causedby changing in the template, thereby adaptively update the template scale. Test thealgorithm results under the framework of Meanshift. The experiment proved that,SIFT features can effectively solve the problem.Finally, in-depth study of the characteristics and effects of sparse appearancemodels for extended target tracking algorithm performance. From the perspective ofthe target appearance model describes the expansion of target tracking. Real-Timecompressive tracking was a simple and effective tracking algorithm. However, therewere a number of problems which need to be addressed. First of all, it was easy tointroduce errors due to factors such as occlusion and clutter. Secondly, it couldn’tupdate the positive and negative samples accurately while using fixed trackingwindow. At last, the number of testing samples was too large, which affected thespeed of tracking. The occlusion was checked by comparison between consecutiveframes’ histograms, and the coefficient can be also updated adaptively by thecomparison result. We searched for more specified areas with multi-scales to find outthe best matching place, and to handle scale change of the target on the basis of theoriginal algorithm’s tracking result. The different numbers of sub features sets wereutilized to filter the testing samples. In that case, the speed of tracking process wouldbe improved. The strategies we proposed would improve the original algorithm’sperformance to avoid the failure of tracking. The experimental results indicate thatthe algorithm can run in real-time and perform favorably against state-of-the-artalgorithms on challenging sequences in terms of efficiency, accuracy and robustness.Further more, In this paper, we propose an algorithm based on SIFT andcompressive features to develop effective and efficient appearance models for robustobject tracking due to factors such as pose variation, illumination change, occlusion, and motion blur. The algrithm describe the target and background with compressivefeatures which labled as positive and negative specimens sampling from frames. Thetracking task is formulated as a binary classification via a SVM classifier with onlineupdate in the compressed domain. In new frame, utilize the classifier to abtain thetarger’s position. Meanwhile, introduce SIFT to solve the target size change, so as toachieve adaptive template size. The proposed tracking algorithm performs favorablyagainst state-of-the-art algorithms on challenging sequences in terms of efficiency,accuracy and robustness.更多还原