【作者】 夏侯玉娇；

【导师】 龚声蓉；

【作者基本信息】 苏州大学 ， 计算机应用技术， 2009， 硕士

【摘要】 视频监控技术是计算机视觉领域一个新兴的应用方向和备受关注的前沿课题,是计算机科学、机器视觉、图像工程、模式识别和人工智能等多种学科的结晶,广泛应用于城市道路交通监控,安防监控等各方面。运动对象分割是视频监控中的一项关键技术,分割的准确性直接影响后续任务的有效性,因此具有十分重要的意义。目前,对视频对象的分割有很多优秀的算法,但这些算法都是针对特定的应用提出来的,而适合任何场景的全自动视频分割算法仍然是一个有待解决的经典难题。本文基于LK光流法和混合高斯模型的视频对象分割算法进行了深入研究和大量实验,得到了一系列有价值的结论和研究成果。本文主要工作可总结为以下几个方面:1.分析现有光流法的性能及优缺点,提出了一种改进的算法。本文运用Gaussian金字塔降低运动对象的速度,使用结合Gaussian分布的LK光流法进行分割,得到运动对象。该算法提高了实时性,能获得更加精确的运动对象。2.考虑到初始背景帧的提取好坏直接影响到背景建模的性能,在现有的初始背景帧的提取方法基础上,提出了一种MEAMO方法来产生初始的背景帧,减少了初始误差,有利于后续任务的完成。3.由于阴影对运动对象的分割会产生严重的影响,提出了一种基于混合高斯模型的自适应阴影检测算法。该算法选择在CIE LUV颜色空间,利用亮度分量L对背景建模,再利用高斯分布对前景和背景中的L分量比值进行自适应阴影检测。该算法实现了自适应的阴影检测,具有较强的鲁棒性和较高的分割精度。总之,本文对视频监控中的分割算法做了进一步的研究,通过实验证明,取得了满意的实验结果。更多还原

【Abstract】 Video surveillance technology is the new-emerging application direction in the filed of Computer Vision and is attracting more and more attention.It spans many subjects including computer science,machine vision,image engineering,pattern recognition, artificial intelligence and so on.It widely applies in city road traffic surveillance and security surveillance.The movement objects segmentation is the key technology in video surveillance.The veracity of segmentation directly affects the effectiveness of latter tasks.So it has very important meaning.At present,there have many excellent segmentation algorithms for video objects,but these algorithms are proposed in special application environment.The automatic video segmentation algorithm that can use in any scene is still a classical difficulty need to be solved.The video objects segmentation algorithms are deeply studied based on LK Optical Flow and Mixture Gaussian Model in this thesis.Through a great deal of experiments,the thesis acquired a series of valuable results which can be summarized in the following aspects:1.On the basis of analyzing the performance of existing Optical Flow methods,an improved algorithm is presented.This thesis uses Gaussian pyramid reduce the speed of movement objects,and then segments pictures use Gaussian distribution combine with LK Optical Flow.The algorithm improves the real-time capability and can get more accurate movement objects.2.Considering the extraction effects of initial background frame directly impact the performance of background modeling,an MEAMO algorithm of producing initial background frame is proposed based on existing methods.The algorithm can decrease the initial error and it is helpful to achieve the latter tasks. 3.For the shadows produce serious impacts to movement objects segmentation,an adaptive shadows detection algorithm based on Mixture Gaussian Model is presented. Choosing CIE LUV color space,the paper uses L weight do background modeling,and then uses Gaussian distribution do adaptive shadows detection for the ratio of foreground’s L weight and background’s L weight.The algorithm achieves adaptive shadows detection.It has strong robustness and high accuracy.In conclusion,this thesis makes farther research for the segmentation algorithms in video surveillance.The experimental results of each researched algorithm are intending and good.更多还原