基于特征编组的目标跟踪方法研究
文献类型：学位论文
| 作者 | 邵春艳
|
| 学位类别 | 博士 |
| 答辩日期 | 2016-05-26 |
| 授予单位 | 中国科学院沈阳自动化研究所 |
| 导师 | 丁庆海 ; 罗海波 |
| 关键词 | 目标跟踪 仿射不变特征提取 边缘提取 直线段提取 特征编组 仿射一致性 相似性度量 |
| 其他题名 | Research on Target Tracking using Feature Grouping |
| 学位专业 | 模式识别与智能系统 |
| 中文摘要 | 基于模板匹配的目标跟踪算法的关键技术手段主要包括两个部分：目标特征的选择以及跟踪算法的相似性度量函数的设计。因此，本文将基于模板匹配的目标跟踪算法的设计问题归纳为两个主要的难点问题：1.如何提取稳定的图像特征，本文主要针对图像底层特征提取算法进行研究，使得所提取的特征能够最大限度地适应目标运动过程中外形的形变、光照的变化以及其它因素的影响；2.如何鲁棒地度量目标模板与实时图像上侯选区域的相似程度，从而有效地在后续图像帧中定位目标。本文的研究工作主要围绕上述两个难点问题进行展开，完成了以下四个方面的研究工作：1.首先，根据卫星遥感与飞机航拍系统获取的图像中目标变化满足仿射变换这一特性，将本文中的图像特征提取研究工作定位在仿射不变性特征提取的研究上。对当前国内外在仿射不变特征上的研究工作进行归纳总结，针对全局仿射不变特征提取的典型算法——MSA(Multi-Scale Autoconvolution)变换提取的仿射不变量对光照变换敏感这一问题，进行了改进，提出一种改进的MSA变换，同时提出一种N-邻域向量夹角映射概念，以像素点间的N-邻域向量夹角映射代替传统MSA变换中概率密度函数，从而避免了传统MSA变换中采用像素灰度作为概率密度函数对光照变化比较敏感的问题。实验以不同光照条件下采集的图像为测试图像，在这些图像上提取的仿射不变量验证了改进的MSA变换能够很好地适应光照变化，与MSA变换以及仿射不变矩两种仿射不变量提取算法的对比实验验证了改进后的MSA变换能够提取稳定的仿射不变量，同时对目标的尺度、遮挡等干扰的适应性较好；2.其次，考虑到图像中共线点的线性关系具有仿射不变性，本文对图像的线型特征开展了研究。主要开展了图像边缘检测算法与图像直线段检测算法两个方面的研究工作，对目前典型的边缘检测算法进行总结，重点对Canny算法的计算过程进行介绍并分析了算法中存在的问题，如非极大值抑制过程导致提取的边缘结果中存在边缘抑制现象，以及阈值选择过程中阈值大小选择问题等。针对这些问题，提出一种结合角点检测的边缘检测算法，首先根据角点确定图像中边缘存在的侯选区域，将全局处理问题转化为局部化处理，避免了进行边缘检测时全局阈值设置环节；然后在这些区域内采用改进的RATMIC描述子进行边缘线段检测，这种改进的RATMIC描述子能够很好的适应非线性光照变化，同时避免了边缘阴影与杂波现象。与近年来比较典型的边缘检测算法的比较实验验证了本文提出的边缘检测算法对边缘细节具有较好的描述性，同时能够适应目标的仿射变换以及图像的光照变化。在边缘检测算法的研究基础上，结合小特征值分析，提出一种基于小特征值分析的直线段检测算法，能够解决原始直线段检测算法在光照条件变化时检测结果不稳定的问题，在保证检测直线段的数量的同时，提高了检测直线段的连续性与覆盖率。3.然后，由于本文后续跟踪过程中提出一种新的相似性度量算法，在完成对模板与跟踪帧的特征提取工作之后，需要将模板与实时图上的特征点匹配，本文提出一种基于空间纹理相似性的角点特征匹配算法。该算法以Harris角点作为匹配特征，通过计算目标角点空间距离矩阵在角点邻域纹理特征向量上的瑞利商，将目标结构的显著性进行量化分析，通过比较两帧图像上角点特征的显著性对这些角点进行匹配，与NCC(Normalized Cross-Correlation)算法的比较实验验证了算法的高效性。但是由于算法的实时性仍然较差以及匹配率略低，为了提高特征匹配算法的运行速度同时提高算法后期相似性度量算法的精度，本文对目标的特征点集进行编组，以编组作为匹配算法的数据处理单位进行匹配。引入心理学家提出的“格式塔规则”中的“邻近性规则”与“相似性规则”，根据目标上的特征点的空间分布具有不同坐标方向的密度一致性，提出一种实时的基于空间邻近规则的单帧图像特征点编组算法，能够将目标几何结构邻近的特征点编组为同一特征编组。由于本文算法是在目标发生仿射变换下进行研究的，引入“相似性规则”，提出一种基于格式塔规则的双帧图像特征编组匹配算法，根据相邻图像上特征编组间具有的仿射相似性，将相邻帧中目标对应的特征编组有效地匹配。实验将基于格式塔规则的双帧图像编组匹配算法与基于空间纹理相似性的特征点匹配算法进行比较，得到的匹配正确率与匹配时间表明，以特征编组为单位对图像进行匹配不仅能够降低匹配算法的运行时间，同时能够提高匹配正确率。4.最后，提出一种基于仿射一致性的目标相似性度量算法，通过数学公式推导证明了该方法可以应用于刚体目标的跟踪中，为了解决传统相似性度量算法对于特征敏感等问题，将传统的目标特征间的相似性度量问题转换为变换空间内的度量问题进行处理。由于本文进行相似性度量的数据为仿射单映阵，仿射单映阵集在空间中构成李群集合，首先采用李群聚类对仿射单映阵集进行聚类，算法能够较好地将空间相似的那些仿射单映阵聚类，从而实现仿射单映阵的度量，但是由于算法在实际应用中的实时性与精度较差，本文提出一种基于高维空间数据聚类的仿射单映阵度量算法，根据Hartigan定义的密度聚类函数设计了一种聚类密度函数，能够很好地将六维空间内相似的仿射单映阵聚集在一起。最后将这种基于高维空间数据聚类的仿射单映阵的度量算法应用于目标跟踪中，在飞机序列上的跟踪实验表明，该算法能够较好地对发生仿射形变的目标进行跟踪，同时能够很好地适应背景的变化。与其它经典目标跟踪算法的比较实验也验证了本文跟踪算法不仅能够稳定地跟踪目标，而且对目标的遮挡、尺度等均具有较好的适应性。 最后一章对课题所做的研究工作进行了归纳总结，并给出本文的主要创新点以及算法目前仍未解决的问题，同时基于本文所做的研究工作，探讨了未来目标跟踪研究领域仍需开展的研究内容。 |
| 英文摘要 | Two main factors that impact the template matching tracker are the image features and the similarity measure. Therefore, this thesis concluded the tracking issue into two focuses to research. One is how to extract the stable image features that can be invariant to the image illumination, shape changing and other variation. The other one is how to measure the similarity between the template and the candidate target region in the tracking frame in order to track effectively in the following frames. This thesis concentrates onto the two mentioned focuses to explore the following research topics in target tracking. First, as the transformation between different remote sensing images and aerial images is affine transformation or similarity affine transformation, this work has researched the affine invariant feature extracting algorithms. Several state-of-the-art algorithms on the affine invariant feature extracting are surveyed. An improved MSA transformed is proposed aiming at to addressing the problems of MSA transform on the sensibility to illumination variation. We propose a novel N-domain vectors included angle map to adapt the illumination variation. N-domain vectors included angle map is calculated through computing the vectors included angle where the vectors are composed of the image point and its N-domain image points. We have proved this method in mathematical aspects that taking the N-domain vector included angle as the probability of the pixel is illumination invariant. This paper illustrates the performance of the improved MSA transform in various object classification tasks. As shown by a comparison with the original MSA transformation descriptors and affine invariant moments, the proposed method appears to be better to cope with illumination variation, image occlusion and image noise. Second, we have researched the image linear features due to that the linear relationship of the pixels in the same line is affine invariant. This thesis mainly focuses on the edge detection and line detection. We introduce the classical edge detection algorithms, and analyze the popular edge detection algorithm-Canny detector. Aiming at addressing the problems of Canny detector such as non-maximum suppression and selection of the threshold, etc, we have proposed a novel edge detection algorithm-edge detection using corner detection. First, to transform the global problem into the local problem, the corner detector is utilized to get the candidate edge areas without setting threshold in edge detection. Second, an improved RATMIC detector that is capable of handling non-linear uniform intensity changes is carried out in these ROIs. Note that the improved RATMIC detector could remove the edge shade and noise. Experiments illustrate the proposed edge detector can extract more detailed edge in the image when illumination changes abruptly compared with other edge detectors. Additionally, a new line detection method is proposed based on the previous edge detection. The proposed line detection can adapt illumination variation and extract more and consistent line segments. Third, a new similarity measure for target tracking is proposed. The features between the template and tracking frame after extracting the features are matched before similarity measure. A new feature matching algorithm using the similarity of spatial texture is proposed. The matching algorithm uses Harris corner as matching feature. We calculate the spatial distance matrix of the corners in the image objects, and transform the measure of the image corners into the measure of spatial texture amplitudes by calculating the Rayleigh quotient of the spatial distance matrix in the LBP feature space. Corners in different images were matched through comparing their corresponding Rayleigh quotients. The proposed feature matching algorithm is more effective than matching based on NCC(Normalized Cross-Correlation). However, the computation complex is still difficult to meet real-time processing. As a consequence, the thesis groups these features to decrease the calculation time. As analyzing the similarity and the proximity in the Gestalt laws, a real-time feature grouping algorithm based on the spatial distribution is proposed to group features in a single frame. The grouping algorithm can group the spatial features according to their density consistency in different plane coordinates. Moreover, a feature groups matching method is proposed using the affine invariance. The matching algorithm can match the feature groups between consecutive frames effectively according their affine similarity. Comparison between the feature groups matching algorithm and corners matching algorithm has been carried out on the plane aerial images. The matching rate and computational time demonstrate that matching based on feature groups can not only decrease the running time, but also improve the matching rate. At last, this thesis proposes a novel similarity measure based on the affine consistency of rigid object. Mathematically, we prove the similarity measure is available on rigid object tracking. To overcome the sensitivity of traditional similarity measure on feature extracting, the similarity measuring in the feature space is transformed into that in the transformed space. As the affine transformed matrixes compose a Li group, we use the Li-group clustering to measure these affine transformed matrixes. However, the clustering applied on the experimental data is time-consuming with low accuracy. Therefore, we propose a similarity measure on affine matrixes using high-dimension data clustering. According to the density function proposed by Hartigan, we construct a new clustering density function that can cluster the affine matrixes in 6-D space effectively. Finally, the high-dimension data clustering is utilized into the target tracking. The extensive experimental results demonstrate the proposed tracker can track object with affine deformation. Our tracker produces stable tracking results under occlusion and scale changes. The last chapter gives the conclusion of this thesis and details the innovation. Several problems that are still existed in our work are discussed to decide the future research interests. We list the future exploring focuses according to these problems. |
| 语种 | 中文 |
| 产权排序 | 1 |
| 页码 | 143页 |
| 源URL | [http://ir.sia.cn/handle/173321/19624]  |
| 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 | 邵春艳. 基于特征编组的目标跟踪方法研究[D]. 中国科学院沈阳自动化研究所. 2016. |
入库方式： OAI收割
来源：沈阳自动化研究所
浏览0
下载0
收藏0
其他版本
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。