For　the　accuracy　and　real-time　requirements　of　stereo　matching　when　mobile　robot　recognizes　the　target,　an　improved　winner-take-all　(WTA)　algorithm　which　is　based　on　the　parallel　binocular　system　is　proposed.　First,　extract　the　relatively　big　difference　points　between　two　images　as　feature　points.　Then,　implement　the　stereo　matching　for　the　feature　points　using　WTA　algorithm　and　only　implement　a　simple　verification　for　nonfeature　points.　Nonfeature　points＇　disparity　values　arc　the　ones　of　neighboring　pixel.　Finally,　obtain　a　dense　disparity　map.　The　extracted　feature　points　focus　on　the　disparity　discontinuity　regions　and　the　matching　accuracy　of　the　algorithm　is　equivalent　to　other　existing　algorithms.　But　calculation　speed　of　the　algorithm　is　faster　and　its　edge　feature　is　better.　So　it　is　a　stereo　matching　algorithm　with　accurate　matching　and　good　real-time.