在多高斯模型的基础上，从场景中模型分布不均匀性出发，提出了一种新的快速背景差算法。该算法针对混合高斯模型中固定模型数量不足的问题，建立了模型产生和退出的机制，使模型数量能够自动适应场景特点，实现了高斯模型的实时自适应分布，即提高了准确性又有效地减少了模型的总量；同时，针对混合高斯模型中计算量大的问题，对模型参数的计算进行了优化，将耗时的浮点运算转化为整型运算，减少了计算量；算法中引入了生存时间和模型重现频率的概念，通过对模型重现频率的限制有效抑制高频噪声。与混合高斯模型的实验结果对比说明，该快速算法保持了原算法的优点，执行速度提高1倍以上，检测结果准确，算法内存消耗小，前景轮廓清晰，抑制高频噪声的能力强，整体效果优于混合高斯模型的背景差算法。

This paper presents an efficient background subtraction algorithm using multiple scene models to cope with variations of noises in a background. A mechanism has been developed to add and delete scene models so that the distribution of the models is adaptive to the background characteristics. The calculation for the model parameters has been optimized so as to avoid time-consuming floating point calculation. We introduced the living time and recurrent frequency to the models so that the algorithm can suppress high frequency background noises effectively by controlling the model recurrent frequency. Experiments using video data have been conducted to compare the performance of our algorithm with that of the mixture Gaussian model algorithm. The experimental results demonstrated that our algorithm can extract the foreground contour more precisely, efficiently and with less memory, while maintaining the advantages of the mixture Gaussian model algorithm. It was also found that high frequency noises that cannot be rejected by the mixture Gaussian model can be suppressed.