Latest　researches　in　stereo　matching　mainly　focus　on　disparity　optimization.　Particularly,　much　effort　has　been　put　into　the　cost　aggregation　to　improve　the　cost　computation　accuracy.　However,　the　performance　of　the　cost　aggregation　is　determined　by　both　the　aggregation　method　and　the　prior　cost　volume　obtained　by　the　matching　cost　computation　based　on　certain　image　features.　Different　choices　of　features　could　result　in　different　quality　cost　volumes.　Therefore,　it　is　necessary　to　study　feature　selection　for　stereo　matching.　Our　experiments　show　that　in　the　region　close　to　the　edges　or　with　high　textures,　features　captured　in　smaller　windows　could　lead　to　a　better　performance.　Based　on　this　observation,　a　novel　gradient　image　guided　feature　selection　method　is　developed,　which　takes　the　gradient　image　of　the　left　frame　as　a　reference　image　to　regulate　the　image　feature　selection.　Experiments　based　on　Middlebury　benchmark　have　verified　the　superiority　of　the　proposed　algorithm.　©　2015　Taylor　&　Francis　Group,　London.