【作者】 李炜；

【导师】 吴晓娟； 刘允才；

【作者基本信息】 山东大学 ， 通信与信息系统， 2011， 博士

【摘要】 群体目标动态分析是计算机视觉和数字图像处理领域的前沿课题,属于运动目标分析范畴。该课题针对各种不同运动群体进行背景消除、运动流分割、群体密度估计、群体跟踪和群体行为理解等研究。群体动态分析技术的应用十分广阔,包括智能监控、安全检查和海洋生物检测等等。现在,该技术已经在某些方面给人类的生活研究带来便利,随着该技术的发展,其必将更加深远的造福人类。因此群体动态分析的研究具有很高的科学价值和深远的实际意义。目前运动目标分析领域中,对单个或多个运动目标的检测分析已达到较高水平,包括目标检测、跟踪、识别和行为理解等技术,无论从理论还是实际应用都已成熟。而群体运动目标分析研究中存在许多难以解决的问题,限制了该课题的发展。群体运动,相对于单个或多个目标运动,具有运动环境复杂、目标数量众多、运动快慢不一、目标之间存在遮挡等特点,给课题研究带来很大困难。因此,国内外学者们对此课题的研究较少,上世纪末才有人涉足该课题研究,最近几年才出现一些阶段性研究成果。本文在群体目标动态分析研究中,提出以下创新点：(1)提出全局光流和高斯背景建模融合的背景消除算法,以获取理想的前景。精确的前景提取是提高群体动态分析准确率的前提。全局光流法和混合高斯模型是群体前景检测效果较好的两种方法,但仍不能很好地去除背景噪声。在此基础上,本文提出对此两种方法并行处理,根据其噪声位置和形状的不同,通过融合光流法和高斯建模法的阈值化前景,消除了几乎全部背景干扰。与现有算法相比,该融合算法准确率明显提高。(2)提出光流场角度聚类算法,以实现前景运动流分割。运动流分割是群体动态分析研究中非常新的技术,研究规则运动群体中不同运动流的分割。由于全局光流场包含所有运动像素的速度和方向,因此通过对光流场角度的分割,能够区分群体前景中不同方向的运动流。本文提出对前景光流场角度聚类的方法,可分割双向、三向、U形和环形等不同类型的运动群体。相比现有的流体力学分割和谱聚类分割,本算法的准确率更高。(3)本文提出一种块吸收算法,以消除运动流分割后出现的杂质块。运动流分割中以不同颜色表示不同类别,而杂质块是某种颜色区域中的异色像素块。利用该特性,通过自适应改变杂质块的颜色,可实现快速准确的块吸收。(4)提出基于光流角度直方图的前景运动流分割算法。由于光流角度聚类法耗时较长,本文还提出一种更快速的分割算法。首先将前景光流场角度信息转化成灰度图像,然后从灰度直方图曲线中检测出极大值点和极小值点,随后将表示方向相近的极值点合并,获取一组最终的极小值点。最后,以该组极小值点为分割点实现前景运动流分割。相比于聚类方法,该算法无需循环聚类,处理速度明显提高。此外,分割点选择的准确度高,有效抑制了杂质块的出现。(5)针对群体目标密度估计提出一种多特征分析方法,密度估计与分类是分析运动群体的最有效手段。由于现有的单一特征估计法准确率不高,本文提出多特征估计,从每条运动流中提取四种特征量,其中前景面积和边缘长度从像素层面反映了群体密度,零阶正交矩和灰度共生矩阵的纹理特征从更高级层面反映群体密度。然后采用多元线性回归训练特征向量与密度等级的关系,并确定各个特征量在密度估计中的权重,最终实现群体密度估计与分类。相比于其他只采用一种特征量的密度估计的方法,本算法估计准确率提高了约十个百分点,而且本算法无需将图像分割成许多小块处理,所以算法处理速度明显提高。(6)针对密集群体密度估计,提出一种基于万有引力边缘检测的新算法。密集群体中存在大量遮挡,给群体分析造成巨大困难。现在仍然缺少解决遮挡问题的方法,本文在研究中找到了一个突破口。本算法以万有引力理论为基础,计算前景中像素点之间的引力值作为图像边缘。本算法能检测密集群体中的不明显边缘,通过数字图像处理手段,从二值化边缘中提取出网络边缘结构。该结构的复杂性与群体密度成正比。利用网络边缘中的网格数量和边缘长度作为特征,估计密集和高密集群体密度。实验结果显示,相比于现有算法,本算法估计准确率提高约五个百分点,同时算法处理速度更快。更多还原

【Abstract】 Crowded objects dynamic analysis (CODA) is an advanced subject in the field of computer vision and digital image processing, it belongs to the area of moving targets analysis, it contains the techniques of background noise removal, foreground flows segmentation, crowded objects density estimation, objects tracking and objects behavior understanding. The technology of CODA has broad application prospects, including intelligent surveillance, safety inspection, ocean biological detection and so on. Now, humans have already benefited from CODA, with the development of this technology, it will benefit we people more deeply and far. Therefore, the research of CODA has high scientific value and far-reaching practical significance.Currently, in the area of moving objects analysis, the research of detecting and analyzing single or multiple objects has reached a high level, including objects detection, tracking, recognition and behavior understanding, both the theory and the practical applications are mature. Compared to single or multiple moving objects, crowded objects movement has more complex scenes, more objects, occlusions among objects and so on, these cause great difficulties to CODA. Therefore, few domestic and foreign scholars have studied on this subject, in the end of last century, there were a few of researchers started the research, until now there are some initial achievements.In the research of CODA, we proposed the following innovations:(1) Proposing a combination approach of global optical flow and Gaussian background model, in order to obtain more exact foreground. The exact foreground is the base of high precise CODA. Global optical flow and Gaussian background mode are best choices in foreground detection, but both of them can not eliminate all the noises. Based on this, this paper proposes parallel processing of both the methods. Since the different appearing places of the noises in the two binary foregrounds, by combining the two methods almost all noises can be eliminated. Compared to existed methods, this combination approach is more precise. (2) Proposing an optical flow angle clustering approach, to segment different flows in the foreground. Flows segmentation is the key technology in CODA. Because the optical flow field includes the spped and direction informations of all moving pixels, so by segmenting the optical flow angles, the foreground flows can be consequently separated. This paper proposed an optical flow clustering approach, which can be used to segment double-direction, three-direciton, U and circular flows. Compared to fluid dynamics segmentation and spectral clustering, the proposed approach is more efficient.(3) Proposing a block absorbing approach to eliminate the small blocks appearing after flow segmentation. Since different colors are used to represent different flows, so by adaptively changing the color of small blocks, all of them can be absorbed fast and exactly.(4) Proposing another flow segmentation based on optical flow angle histogram. Since the optical flow clustering approach is quite time-consuming, we propose a fast method as an improvement. First, the foreground optical flow angle is changed to a gray-level image, then we can get the derivative histogram curve of this gray image. Next, checking the zero-value points on this curve to find out maximum and minimum points of histogram. Finally, the updated minimum values are utilized as segment points to devide different flows in foreground. Compared to optical flow clustering, this approach is much faster and has better performance on flows segmengtation.(5) Proposing a multi-feature approach for crowded objects density estimation. Density estimation and classification is the efficient approach in CODA. Since the existing single feature estimation methods have lower accuracy rate, thie paper proposed to extract four kinds of feature for estimation. These features imply crowded objects density from different aspects. Then using multiple linear regression to train the relationship between multi-features and density levels. Finally, this approach achieves density classification. Compared to single feature estimation, our approach has better performance on estimation accuracy. Meanwhile, the time-consuming of this approach is shorter than texture and moment analysis methods.(6) For over-crowded objects, a gravitational edge detection approach is proposed for density estimation. In crowded objects, the occlusion is a very difficult problem to slove, this paper proposes a new solution. This approach is based on the law of universal gravity, it calculates the gravitation between pixels as the edge. This approach can be used to detect the unconspicuous edge in crowded objects, then obtain the edge network. The number of grid and length of the edge network are utilized as features to estimate crowded objects densities. Experimental results show, compared to existing methods, this approach has higher accuracy rate on dense and very dense objects densities estimation.更多还原