Extracting　robust　video　feature　has　always　been　a　challenge　in　the　field　of　video　classification.　Although　existing　researches　on　video　feature　extraction　have　been　active　and　extensive,　the　classification　results　based　on　traditional　video　feature　are　always　neither　flexible　nor　satisfactory　enough.　Recently,　deep　learning　has　shown　an　excellent　performance　in　video　feature　extraction.　In　this　paper,　we　improve　a　deep　learning　architecture　called　ELU-3DCNN　to　extract　deep　video　feature　for　video　classification.　Firstly,　ELU-3DCNN　is　trained　with　exponential　linear　units　(ELUs).　Then　a　video　is　split　into　16-frame　clips　with　8-frame　overlaps　between　consecutive　clips.　These　clips　are　passed　to　ELU-3DCNN　to　extract　fc7　activations,　which　are　further　averaged　and　normalized　to　form　a　4096-dim　video　feature.　Experimental　results　on　UCF-101　dataset　show　that　ELU-3DCNN　can　improve　the　performance　of　video　classification　compared　with　the　state-of-the-art　video　feature　extraction　methods.　©　Springer　Nature　Singapore　Pte　Ltd.　2018.