Fires　often　pose　a　major　obstacle　to　process　safety.　How　to　detect　fire　earlier　and　obtain　more　detailed　fire　hazard　information　has　become　an　important　issue.　Existing　means　of　fire　detection　suffer　from　problems　such　as　insufficiently　detailed　information.　Therefore,　with　the　combination　of　binocular　stereo　vision　and　U-net,　this　paper　proposes　a　parallel　network　structure　model　to　realize　three-dimensional　spatial　localization　and　size　measurement　of　flame.　Two　serials　of　experiments　were　designed　for　different　fire　scenarios　covering　solid　fires　(candle)　and　liquid　fires　(n-heptane)　from　different　scales　that　may　occur　in　city　operations.　In　terms　of　flame　pattern　segmentation　results,　the　precision　of　the　method　can　be　basically　maintained　at　over　90　%.　In　terms　of　three-dimensional　spatial　localization　of　the　flame,　the　MRE　(Mean　Relative　Error)　for　both　type　of　fires　tested　are　mostly　less　than　15　%　for　a　variety　of　parameters.　In　terms　of　size　measurement　of　the　flame,　the　MRE　for　both　the　height　and　the　width　of　the　flame　was　less　than　18　%.　The　implementation　of　this　method　facilitates　the　acquisition　of　detailed　quantitative　data　concerning　flame,　thereby　aiding　in　the　preservation　of　industrial　safety　and　environmental　protection　during　the　occurrence　of　fires.