Background subtraction method through calculate the difference between the current frame and the background model to achieve the detection of moving targets, so the background modeling is a key to the method. Gaussian mixture model can model on the existence of a gradient and repetitive motion scene, and it effectively improve the accuracy of the modeling under the complex scenes of the light intensity changes and objects swing. But it also has its inherent drawbacks, In this paper，for?the?defect of easy to fall into the local optimum of the solution space when using traditional EM algorithm to estimate GMM parameters, we use the the maximum punishment of the likelihood function to resolve its deficiencies; in addition, we advance a differential evolution algorithm based GMM parameter estimation method when do not need to meet the real-time in the detection; Finally, we put the improved GMM parameter estimation method to applied to the moving target detection based on GMM model, and get a good result.