In　large　space　structures,　the　latest　fire　detection　methods　are　based　on　video　image　processing　and　data　fusion.　But　the　false　positive　rate　and　false　negative　rate　remain　unsatisfactory　and　need　improving.　The　emphases　of　this　paper　are　target　extraction　and　recognition.　A　new　adaptively　updating　target　extraction　algorithm　(NAUTEA)　is　proposed　by　which　the　intact　target　can　be　extracted　in　time.　In　addition,　some　fire　video　image　recognition　algorithms,　such　as　fuzzy　neural　network　(FNN)　and　FGALSSVM　(Fuzzy　GALSSVM),　are　studied　and　improved.　To　verify　the　performance　of　these　algorithms,　a　prototype　system　is　developed,　and　a　series　of　algorithm　tests　on　a　fire　video　are　conducted.　These　tests　make　it　clear　that,　the　accurate,　robust　and　real-time　fire　detection　can　be　realized.