【作者】 李莉；

【导师】 齐美彬； 蒋建国；

【作者基本信息】 合肥工业大学 ， 信号与信息处理， 2009， 硕士

【摘要】 运动目标检测作为计算机视觉领域的一个重要分支与基础,在军事、交通、工业以及生物医学等领域具有广泛的应用前景,一直受到广泛的关注,并成为计算机视觉领域的一个研究热点。但是由于运动目标检测问题本身的复杂性,运动目标的检测依然面临着诸多挑战。本文在现有研究成果的基础上,对静态与动态场景下的运动目标检测进行了深入研究。对于静态场景下的运动目标检测,本文提出了一种基于码本的改进型算法,并用多组室外、室内测试视频序列进行了测试验证,测试结果表明其完全能满足应用需求。其中的主要工作及创新包括以下几个方面:(1)基于新的颜色模型,提出了新的高亮与阴影判决准则;(2)本算法中应用了新的码本提炼、码本更新及自学习策略;针对动态场景中的运动目标检测,本文采用了两种算法进行运动估计与补偿。在第一种算法中,采用了基于块匹配的全局运动估计算法,该算法中采用特征块选择模板、Canny边缘检测、新的三角形菱形搜索算法及运动矢量提炼等思想,来降低算法的运算复杂度,并确保与提高运动估计的精确度。但由于该算法采用了平移模型,因此仅适用于全局运动为平移模式的视频序列。在第二种算法中,采用了基于特征匹配的全局运动估计算法。为了准确估计运动模型的参数,采用了特征点选择模板、SIFT算子、最小二乘法、特征点匹配对校正及双线性插值相结合的方式。由于该算法采用了六参数仿射模型,因此,可以适用于移动、旋转、缩放等全局运动模式,从而具有普适性。二种算法都采用了一组或多组室外与室内标准视频序列进行测试验证。实验结果表明,两种算法在其各自的应用范围内具有较好的鲁棒性和有效性。这部分内容包括的主要创新工作如下:(1)提出了一种提炼运动矢量的阈值法;(2)提出了一种新的三角形菱形搜索算法;(3)提出了一种新的特征点匹配对校正策略。更多还原

【Abstract】 Moving object detection is an important branch and foundation of computer vision; it has extensive application prospects in many fields, such as, military applications, traffic, industries and bio-medical, et al, and attracts broad attention, and becomes a hot research topic in computer vision field. Because of its inherent complexity, moving object detection still faces lots of challenges. In the paper, based on the existing research achievements, the methods of moving object detection in static and dynamic scene are studied deeply.For moving object detection in static scene, the paper proposes a modified codebook algorithm, and performs validation experiment with several indoor and outdoor video sequences. The results prove that the proposed algorithm can meet with the application requirements. The original work in this section mainly includes the following points:(1) Based on a new color model, a new high-light and shadow judge rule is proposed.(2) A new codebook abstracting, real-time codebook updating, and self-study policy is used.For moving object detection in dynamic scene, the paper introduces two algorithms to perform motion estimation and compensation. Firstly, a global motion estimation algorithm based on block matching is proposed, this algorithm combines several ideas, such as feature selection templet, Canny edge detection operator, new triangle-diamond search policy, and motion-vector abstraction, et al, to reduce the operation complexity, and insure and improve the precision of motion estimation. However, this algorithm uses translation transform model, it is only suitable for video sequences whose global motion style is panning. Then, a global motion estimation algorithm based on feature matching is proposed. In order to estimate the parameter of global motion model well and truly, it combines several ideas, such as feature selection templet, SIFT operator, least squared criterion policy, revision of feature points matching-pairs, and bilinear interpolation, et al. Since this algorithm uses affine transform model, it has universality, and can suitable for video sequences whose global motion style including panning, tilting, zooming, et al. The two algorithms are all validated with one or several standard indoor and outdoor video sequences. The results demonstrate that the two algorithms have good robusticity and validity in respective application fields. The original work in this section mainly includes the following points:(1) A threshold policy is proposed to abstract the global motion vectors.(2) A new triangle-diamond search policy is proposed.(3) A new revising policy of feature point matching-pairs is proposed.更多还原