Abstract: since 　there 　are 　severe 　geometric　distortions 　between　obli que 　imagesit 　is 　difficult 　for 　the 　available 　al gorithms 　to　obtain　satisfactor y　quasi－dense　matchingthus 　a　serious 　obstacle 　to 　subsequentthree－dimensional 　reconstruction．this 　paper 　proposes 　an　al gorithm　of 　automatic　quasi－dense　matchingand　three－dimensional 　reconstruction　based　on　affine 　invariant 　features．the 　proposed　method　can　bedivided 　into 　three 　sta ges seed　re gion　matching．the　affine 　invariant 　features 　are 　first 　extractedand　matchedbased　on　matched　features 　least 　square　matchinglsmis 　implemented．subse quentlythe　maximum　convergence 　re gion mcrof 　lsm　can　be　calculated　by　adaptive 　iteration． quasidense　matching．to　detect 　correspondencesthe　normalized　cross 　correlation nccfor 　a　pers pectivetransformation　model　was　adopted 　to 　the　grid 　points 　in　mcrand　the　weighted　sum　of 　a　squareddifference　metric　for 　a　epipolar 　geometr y　correction　model　was　applied 　to 　the 　rest 　of 　grid 　points．afterwardsoutliers　will　be　eliminated 　based　on　the　geometrical 　consistenc y　check． three－dimension－al 　surface 　reconstruction　of 　the 　3dpoint 　cloud 　commenced　with　the 　application 　of 　a　image　orientationand　collinearit y　equation．thena　3dsurface　model　was　generated　based　on　the 　3dpoint 　cloud．experiments　with　obli que　stereo 　images 　demonstrate 　the 　effectiveness 　of 　the　proposed　al gorithm