Gaussian mixture model is a popular method to model dynamic scenes viewed by a fixed camera. However, we found
that it is not a trivial problem for each component of Gaussian mixture to learn the accurate parameters for complex
pixels. Furthermore, traditional method of Gaussian mixture has to make a tradeoff between system stability and
convergence rate. We developed a mechanism of double-layer Gaussian mixture model for moving object detection from
dynamic scenes, which can improve the convergence rate without compromising the system stability. Additionally,
temporal consistency of variances was taken into account to alleviate camouflage problems in the process of detection.