Gaussian　mixture　model　(GMM)　is　an　effective　way　to　extract　moving　object　from　a　video　sequence.　However,　the　conventional　mixture　Gaussian　method　suffers　from　slow　convergence.　In　this　paper,　a　novel　approach　which　combines　Gaussian　mixture　model　(GMM),　three-frame-difference　and　cropped　frame　technique　is　proposed　to　detect　moving　object.　Firstly,　gray　correlation　based　three-frame-difference　method　is　adopted　to　make　sure　the　motion　region.　Then,　a　cropped　frame　technique　is　presented　to　clip　the　motion　region　of　an　image.　And　the　background　model　is　estimated　only　on　the　motion　region,　which　can　dramatically　reduce　computational　load.　At　the　same　time,　we　propose　a　new　initialization　strategy　based　on　dynamic　grid　and　density　estimation　for　EM　algorithm　to　reduce　the　influence　on　initial　values.　Extensive　experimental　results　demonstrate　that　our　approach　can　obtain　satisfying　performances　for　practical　application.　©　2012　IEEE.