Based　on　time　redundancy　in　video　image　sequences　an　adaptive　SIFT　(Scale-invariant　feature　transform)　algorithm　is　proposed.　According　to　the　latest　three　models＇　outputs　in　global　motion　estimation,　the　algorithm　predicts　overlapping　regions　between　reference　and　current　frames　by　using　Lagrange　parabolic　interpolation,　and　then　extracts　feature　points　in　the　smaller　region　instead　of　the　whole　image.　In　this　way,　it　can　eliminate　a　large　number　of　information　redundancies　to　increase　the　processing　speed　of　each　frame,　improve　the　effectiveness　of　feature　points　and　reduce　the　mismatch.　Experimental　results　show　that　the　improved　algorithm　has　the　features　of　strong　adaptive　ability,　rapidity　and　high　matching　accuracy,　and　it　can　be　applied　to　the　real-time　positioning.