In　this　paper,　we　propose　a　novel　method　with　which　we　can　locate　faces　in　a　video　sequence　based　haar-like　feature,　as　an　extension　to　Viola　and　Jones＇　cascade　learning　method　for　object　detection.　Our　method　is　adopted　to　supply　region　of　interest　of　one　frame　for　adaptive　video　coding　based　on　multidimensional　description　space.　The　detector　is　programmed　by　C　language　so　that　it　will　be　integrated　in　AV　S　coder.　Our　goal　is　to　achieve　lower　false　positive　rate　so　that　detector　will　hardly　detect　any　false　face　if　there　is　no　face　in　one　frame　of　a　video.　To　meet　the　target,　we　make　an　improvement　on　the　aspects　of　how　to　select　feature,　set　sample　weight　and　decrease　threshold　of　weak　classifier　and　strong　classifier.　Experimental　results　show　that　our　detector　yields　high　classification　performance　over　frame-based　detection.　1548-7741/Copyright　©　2008　Binary　Information　Press　October　2008.