【作者】 赵红；

【导师】 张小洪；

【作者基本信息】 重庆大学 ， 计算机软件与理论， 2010， 硕士

【摘要】 运用目标提取算法从连续视频图像中提取并优化运动目标,是运动跟踪、目标识别、视频监控、视频摘要、三维重建等问题的关键环节,由于其广泛的应用及本身的复杂性,运动目标检测已成为当前模式识别和机器视觉的热点问题之一。本文依托重庆市科技攻关项目《嵌入式WEB监控视频摘要生成及其快速浏览关键技术研究和开发》(CSTC2009AC2057),研究动态场景中的运动目标检测,该问题不仅是视频监控的核心,更是系统中后期操作的基础与关键。其目标检测的准确与否决定了头部检测、运动跟踪、视频摘要等各种后续处理的效果,因此成为项目研究中的一项重要课题。但目前流行的运动目标检测算法在处理光照变化、背景扰动、运动目标与背景相似等实际问题时,往往难以取得良好的应用效果。针对上述问题,本文的主要工作和创新点如下:首先对运动目标检测的基本算法进行了分析与实现,针对高斯混合模型和码本模型两种流行的背景相减法的不足提出了改进策略。对于高斯混合背景模型,采用梯度序列代替传统的亮度序列计算模型中的概率估计,提取更加清晰的目标轮廓,同时建立高斯概率密度分布的查找表以降低目标像素点与背景模型的匹配时间,最后引入像素点的空间分布一致性去除错误检测,提高目标检测的准确性。在码本背景模型中,采用YUV空间代替RGB空间,降低码本背景模型的光照敏感性,同时在新的YUV空间采用球状码元模型改进原RGB颜色空间的圆柱码元模型,提高了算法的鲁棒性。其次,本文采用阴影剪除、形态学处理、graph cut、区域连通标号等技术作为后验优化操作,剪除目标阴影、剔除复杂场景的干扰噪声、去除目标边缘毛刺,从而得到更加清晰的运动目标轮廓,为系统后续处理提供了准确的运动目标信息。最后,本文结合面向对象的软件设计方法,使用MATLAB仿真和VC++开发平台对运动目标检测模块进行了设计与实现。并针对运动目标检测的各种算法,制定了一套系统的评估框架,定量分析不同方法的优缺点。实验测试和开发的前期效果证明了本文改进的前景提取算法不仅提高了运动目标检测的准确性,而且具有良好的实时性和鲁棒性。更多还原

【Abstract】 Extracting and optimizing target motion from continuous videos is the key link of movement track, target recognition, video monitoring, video abstract, 3d reconstruction problems and so on. Because of its wide application and the complexity of moving object detection, it becomes one of the hot topics in the current pattern recognition and machine vision’s fields.Based on Chongqing torch-plan projects named the key technology research and development of the video abstract formation and fast browsing from the embedded WEB video (CSTC2009AC2057), this study focuses on detecting moving object of dynamic scene, which is not only the core of video monitoring system, but also the foundation and the key for later processing. The accuracy of target detection decides further treatment effect of the head detection, motion tracking and video abstracting, therefore it becomes an important topic of research project. But currently popular target motion detection algorithms have difficulties in dealing with actual applications such as illumination change, background disturbance and motion objects similar to backgrounds. In the view of the above questions, main work and innovative points of this paper are as follows:Firstly the basic algorithm detecting the moving object is analyzed and implemented, and then strategies are proposed to improve the two popular background subtraction: gaussian mixture model and codebook model. For the gaussian mixture background model, gradient sequence replaces the traditional brightness sequence to calculation the probability of the model and extract clearer target outline. At the same time lookup tables of gaussian probability density distribution is established to reduce the matching time between target pixels and background model. Finally introducing spatial distribution of pixel to remove error detection and improve the accuracy of target detection. In this background model of codebook, YUV color space replaces RGB space and decrease the sensitivity to the illumination of background, while space spherical code element model in new YUV instead of cylindrical code element model in the original RGB color space improves the algorithm’s robustness.Secondly, shadow cutting off, morphology processing, graph cut and regional connected technologies are adopted as posteriori optimizing operation, such as cutting off target shadow, eliminating the disturbed noises of complicated scene and purifying edge burr, thus obtains the clearer target outline providing more accurate target motion information for further treatment system.Finally, the object-oriented software method, MATLAB and vc + + are combined to design and implement platform for moving object detection module. And for various algorithm moving of targets detection, a evaluation system framework is developed, and quantitative analysis of the advantages and disadvantages of different methods are obtained. Experimental test and results prove that this paper our algorithm not only improves the accuracy of targets detection, but also have good robustness and real-time performance.更多还原