【作者】 常亚林；

【导师】 冀小平；

【作者基本信息】 太原理工大学 ， 信号与信息处理， 2010， 硕士

【摘要】 在计算机视觉、视频监控领域,背景模型构建是一项关键技术。背景建模的结果将会对图像序列的运动目标检测、分类、跟踪及行为理解等后续处理工作产生重要影响。背景建模是序列图像分析的基础性工作,也是当今国内外学者研究的热点问题。本文对复杂场景的背景建模进行了较为深入的研究,并着重研究了贝叶斯背景估计算法。本文的主要工作如下:首先,利用混合高斯模型算法分别对静态背景、动态背景进行背景建模。通过对实验结果的分析得出,此算法仅对静态背景或较简单的动态背景有很好的效果;然而在处理室外动态复杂场景时,构建出来的背景效果并不理想。然后,主要针对以上混合高斯模型的不足,引入了贝叶斯理论。通过贝叶斯理论对混合高斯模型进行了多方面的改进。第一,针对高斯模型的收敛问题,应用EM算法对其进行了改进,保证算法的快速收敛。第二,在背景分割阈值选取上,根据影响阈值产生变化的因素,提出了由样本差和均值构造的分割阈值选取方法。针对背景分割图中噪声点较多和前景目标连通性较差的问题,在视频分割图像后处理阶段,分别采取连通域处理、噪声抑制和形态学技术等方法进行处理,降低了目标检测的虚警率。最后,把上面模型估计出来的背景应用到滞留物和偷窃物的检测方面。由贝叶斯模型更新速率不同而得到的两个背景,通过背景差法得到两幅二值化的前景帧。然后再对前景帧中的运动信息分别进行累计获得特征图像,最后通过特征图像把图像分为运动物、背景和可疑目标。整个过程不需要对目标进行跟踪,便于实现。实验结果验证了本文算法的有效性和可靠性。更多还原

【Abstract】 In computer vision and video surveillance field, background modeling is a key technology. The higher level projects such as moving objects detection、classification、tracking and behavior understanding depend on the results of background modeling. Background modeling is one of the fundamental parts for video analysis, and which is currently a hot topic widely researched around the world. This paper focuses on the background modeling of complex scenes, and places emphasis on Bayesian background estimation algorithm. The main contributions of this paper are summarized as follows:First, Gaussian mixture model(GMM) is used to model background images which are static and dynamic scenes, and through the analysis of experiment results, we conclude that GMM can get a good effect only when the background is static or partly dynamic;however,while dealing with complex outdoor scenes, results are not nice.Then, Bayesian network algorithm is introduced to improve GMM in this paper,and improvements are proposed for this method. Firstly, for convergence problem of this model, the EM algorithm is applied to ensure fast convergence of the algorithm. Secondly, a thresholding method based on sample mean and standard deviation is presented, which can class pixels more accurately. At last, some measures have been taken in image post-processing such as noise suppression and Mathematical Morphology technology to remove noise and increase connectivity.Finally, this model is applied to detect abandoned and removed objects. By processing the input video at different frame rates, two backgrounds are constructed: one for short-term and another for long-term. Two binary foreground maps are estimated by comparing the current frame with the backgrounds, and motion statistics are aggregated in a likelihood image by applying a set of foreground maps. Likelihood image is then used to differentiate between the pixels that belong to moving objects, temporarily static regions and scene background. The temporary static regions indicate abandoned items, illegally parked vehicles, objects removed from the scene. And this presented pixel-wise method does not require object tracking and can be performed easily.The experimental results show that the approach proposed in this thesis is effective and reliable.更多还原