Global modeling is adopted in existing Constant False Alarm Rate (CFAR) algorithms, and the same distribution model is used which estimates the background clutter to detect the whole area. But practical ground covers complex types, and different ground area has its most suitable backgrounds model. The used model is not fit in some regional, making higher loss of CFAR, bringing down the test performance. So an algorithm is presented which judged the areas according to the different characteristics of background, such as statistical variance and mean ratio. In this way, CFAR detector could select the distribution model on the basis of the regional type automatically and get the best detection results: that is, choosing Gaussian distribution in an uniform region. Weibull distribution is used to eliminate the influence while in a clutter edge, and G0 distribution is used to eliminate the obstacles targets while in a multiple targets interfering region, avoiding mutual shielding effects of adjacent targets.