In this paper

the optimal estimators of the coefficients of a second-order bivariate polynomial for piecewise grayvalue distributions of image are derived by using the least quadrate multiplication

and the remaning components within eath patch can be attributed to white noise with zero meanvalue.Therefore

the quadratic model is set up

which is the best approacl to piecewise grayvauc distributions.In order to detect changes between consecutive frames

the statistical assumption test model which follows F-distribution with degrees of freedon（n+ 3

2n-12） is detived based on the quadratic model

where n is the number of pixels in detection window. Because singe motion target can not be generally selected from the detected changes if there is noly F-test

the techniques of clustering analysis and filling holes are also designed

which can make the target loss and false alarm appronch to zerr. The experimental results made with the television-and infrared image sequences fully show the validity and compatibility of the technique route and key techniques proposed.