In　this　paper　we　describe　our　approaches　to　the　high-level　feature　extraction　task　in　TRECVID　2009.　Our　semantic　detection　system　is　based　on　6　basic　low-level　features　as　well　as　merged　features.　We　have　experimented　with　several　fusion　strategies,　especially　the　low-level　feature　concatenation　and　Borda　fusion.　Our　experiments　showed　that　the　second　Borda　fusion　could　increase　the　number　of　the　true　shots　returned,　but　not　inevitably　improve　the　mean　inferred　average　precision.　Experiments　also　revealed　that　the　selection　of　various　low-level　features　and　fusion　schemes　is　very　crucial　for　achieving　a　good　system　performance.　The　description　of　our　6　runs　for　the　high-level　feature　detection　is　as　follows.