The　video　retrieval　system　we　developed　for　TRECVID　2012　mainly　involves　the　semantic　indexing　task　which　includes　key　frame　extraction,　low　level　feature　extraction,　classification　and　concept　fusion.　We　extracted　a　new　low　level　feature,　explored　various　classification　and　fusion　schemes.　Four　＇light＇　runs　and　two　2　＇pair＇　runs　we　submitted　are　as　follows:　L_A_FudaSys1:　Fusion　based　on　concept　ontology.　L_A_FudaSys2:　Weighted　fusion　of　SVM　and　KNN　outputs.　L_A_FudaSys3:　Average　fusion　of　KNN　results.　L_A_FudaSys4:　Average　fusion　of　SVM　outputs.　P_A_FudaSys1:　Weighted　fusion　of　KNN　and　SVM　Outputs.　P_A_FudaSys2:　Concept　relation　fusion　of　KNN　and　SVM　outcomes.　In　our　experiments,　we　also　implemented　various　special　detectors　to　detect　screen　text,　black　screen　and　human　face　to　enhance　system　performance.