Flame　detection　based　on　computer　vision　is　one　of　the　key　technologies　of　the　modern　surveillance　system.　However,　how　to　reliably　and　accurately　detect　the　flame　is　still　a　tricky　problem.　In　this　paper,　a　video　flame　detection　method　based　on　two-stream　convolutional　neural　network　combining　spatial　and　temporal　features　is　proposed.　Firstly,　the　suspected　flame　region　is　extracted　from　the　video　by　the　combination　of　motion　feature　detection　and　color　feature　detection.　Next,　and　the　extracted　suspected　region　is　classified　by　the　two-stream　convolutional　neural　network.　Finally,　the　region　whose　classification　result　is　flame　is　output　as　the　final　detection　result.　The　experimental　results　on　the　collected　flame　data　set　show　that　the　proposed　flame　detection　method　can　effectively　improve　the　detection　accuracy.　©　2019　IEEE.