摘要
该文提出一种新的基于混合概率模型视频分割方法。这个方法主要利用两个概率模型:隐马尔可夫模型和概率图模型建立一个混合的贝叶斯网概率模型,对视频输入中背景变化的时间和空间局部相关性(同现性)进行学习。在建立正确模型参数的基础上,贝叶斯信念传播算法根据图像输入预测当前背景状态的后验分布。并根据预测得到的背景状态对输入图像进行分割,实验结果显示方法的有效性和在复杂背景变化下的鲁棒性。
This paper proposes an unified framework for detecting and segmenting the foreground objects in complex scenes involving swaying trees, moving shadows and ocean waves. A mixture model of hidden Markov models and probabilistic graphic models is presented to model the variation of background. In this model, the color of each pixel from backgrounds is regarded as a random variable. We employ the property that background variations at neighboring pixels have strong correlation, also known as ＂co - occurrence＂ to initialize and update the mixture model online. Bayesian belief propagation algorithm allows us to efficiently calculate the maximum of the posterior probability of background with input data in this model. Experimental results for real video demonstrate the effectiveness and robust of our method.
出处
《计算机仿真》
CSCD
2006年第4期192-196,273,共6页
Computer Simulation
关键词
隐马尔可夫模型
概率图模型
同现性
贝叶斯信念传播算法
前景目标的检测和分割
Hidden Markov model （HMM）
Probabilistic graphical model
Co -occurrence
Bayesian belief propagation
Foreground object detection and segmentation
