This　paper　proposes　a　junction　detection　method　in　images　by　searching　for　the　intersection　of　the　linear　ridges　in　gradient　domain,　and　then　presents　a　stereo　matching　method　based　on　this　new　algorithm　of　junction　detection,　in　order　to　meet　the　need　of　3D　reconstruction　system.　The　images　are　uncalibrated　and　can　be　taken　by　different　cameras　or　a　single　camera　at　different　time　instants.　The　proposed　method　adopts　a　strategy　from　rough　matching　to　delicate　matching.　In　the　first　matching　process,　the　probability　of　mismatching　is　decreased　on　the　cost　of　reducing　the　matching　number.　Thus,　only　the　junctions　with　high　strength　are　matched　and　the　direction　angles　of　the　junctions＇　branches　are　used　as　a　matching　constraint.　Then,　the　fundamental　matrix　is　estimated　by　RANSAC　and　normalized　8-point　algorithm.　At　last,　the　parameters　of　the　junction　detection　are　modified　and　the　Harries　corner　detection　is　added　to　get　more　feature　points.　More　matching　points　could　be　obtained　through　the　guide　of　the　fundamental　matrix　in　the　second　matching　process.　The　experiment　results　show　that　this　algorithm　has　good　performance　on　speed,　accuracy　and　could　obtain　more　edge　matching　points.　©　2007　IEEE.