【作者】 王亚飞；

【导师】 印勇；

【作者基本信息】 重庆大学 ， 信号与信息处理， 2009， 硕士

【摘要】 视频序列中的运动目标检测与跟踪是一项涉及视频图像处理、计算机视觉、模式识别以及人工智能等多个领域的研究课题,由于其在商业和军事领域中存在广泛的应用价值,近年来一直是人们研究的重点。然而,因为视频图像本身存在固有的复杂性,所以要提出一种稳健而又准确的算法仍然是一个极具挑战的工作。本文主要针对静态场景情况下的运动目标检测和跟踪技术进行研究。在目标检测方面,首先对目前主要的方法进行了归纳和总结,分析了各自的优缺点;然后对基于背景高斯模型的目标检测算法进行了比较详细的分析和论述,并对其中的一些算法进行了改进:本文中对均值和方差采用不同的更新率,给出了一种用模型的匹配次数确定方差更新系数的算法,解决了传统算法中方差收敛缓慢的问题,增加了模型的稳定性和对环境的适应性。由于传统基于混合高斯模型的检测算法仅仅对图像中的每个像素点建模,而忽略了邻域像素之间的相关性,目标检测的准确度不高。因此,本文提出了一种基于空间邻域相关性的目标检测算法,该算法通过重新定义马尔科夫随机场的势能函数,融入了空间邻域的相关性信息,并由此获得了用于运动目标检测的自适应阈值。实验结果表明,所提出的方法对复杂的场景有较好的适应性,能够得到比较准确的检测结果。在运动目标跟踪方面,为了提高算法的实时性,本文采用了基于卡尔曼滤波器的运动目标跟踪算法,首先用卡尔曼滤波器预测目标的大致区域,然后在预测区域内进行目标匹配。为了适应多目标的情况,采用了连通域标记的方法来提取目标特征,建立目标特征链表。并采用匹配矩阵的概念,将跟踪过程分为新目标出现,目标消失,目标遮挡,目标分离,及目标匹配五种情况,并分别做了分析和研究,给出了算法流图;最后本文对目标的匹配搜索算法做了比较详细的阐述,采用了交叉法来进行目标的匹配搜索,进一步提高了算法的实时性。仿真结果表明,本文采用的算法对多目标的情况有较好的适应性,能够获得比较理想的跟踪结果。更多还原

【Abstract】 Moving objects detection and tracking in video sequences is a task related to video image processing, computer vision, pattern recognition, artificial intelligence and other areas. It is an important research topic all the time in recent several years because of its widely application in business and military fields. However, due to the inherent complexity of the video image, it is still a challenging task for proposing a robust and accurate algorithm for moving objects detection and tracking.This paper focuses on moving objects detection and tracking technique under static background. For the moving objects detection, firstly, the detection methods at present are classified and summarized, and their advantages and disadvantages are analyzed. Then, the objects detection methods based on Gaussian model are analyzed and discussed in detail, and a new algorithm is proposed, in which the mean and variance of the model are used at different update rates. The problem of the variance slow convergence in traditional algorithm is solved and the stability of Gaussian model is increased by the algorithm that uses the number of model match to decide update coefficient of variance. Because traditional Gaussian mixture model only builds up the model for a pixel in video image and neglects the spatial local correlation, the accuracy of moving objects detection is not high. In order to solve this problem, a detection method based on spatial local correlation is proposed. The algorithm combines the spatial local correlation with mixture model by redefining energy function of the Markov random field, and then an adaptive threshold is obtained for moving objects detection. The experimental results show that the proposed method can better adapt to the dynamic scenes, and obtain a more accurate objects detection result.For the moving objects tracking, in order to improve the real-time of the algorithm, the tracking method based on Kalman filter is used in this paper. The algorithm firstly use Kalman filter to forecast the approximate position of moving object, and then the object are matched in the forecast area. In order to adapt to multi-objects, the method based on labeling connected domain is used to obtain the features of object and build the feature chain table. Then, a matching matrix is adopted to divide the tracking process into five situations: new object emergence, object disappearance, object sheltering, object separating and object matching. These situations are respectively analyzed and researched, and the flow charts about the situations are showed. Finally, the algorithm of matching and searching is discussed in detail, and the real-time of the algorithm is improved by using cross algorithm to match and search object. The experimental results show that proposed method can better adapt to multi-object detection, and obtain a more perfect objects tracking result.更多还原