Target detection is the basic work for analyzing the behavior of the cows using video analysis technology. It is difficult to extract the moving cows accurately and realtimely with the existing target detection methods because of the complex background environment. In this study, a series of improvement measures were proposed based on Gaussian mixture model to meet the system requirements. A dynamic background modeling method with penalty factor was proposed for the mathematical model of the background which can overcome the high model complexity. A twoclass classification algorithm based on chromaticity distortion and brightness distortion was proposed to avoid the influence of the shaded area in the foreground detection algorithm. Local update method was proposed to avoid missing the target if it stays for a long time. In order to verify the validity of the algorithm, four evaluation parameters were introduced to test the detection algorithm including model complexity, false detection rate of foreground, false detection rate of background and processing time. Experimental results show that model complexity was 5085% lower than the classical method. False detection rate of foreground and false detection rate of background were 18.18% and 7.52%, which had 19.50 and 13.37 percent lower than the classical Gaussian mixture model. Processing time of average single frame was 29.25% lower. Statistics indicate that the algorithm proposed in this study can improve the detection performance and it is an extension to classical Gaussian mixture model.