Stereo　matching　is　one　of　the　key　technologies　in　the　field　of　computer　vision.　The　depth　map　obtained　by　stereo　matching　contains　the　three-dimensional　information　of　the　scene.　The　use　of　depth　map　is　of　great　significance　in　the　three-dimensional　reconstruction　of　the　map　and　the　autonomous　navigation　of　the　robot.　Aiming　at　the　accuracy　and　speed　of　stereo　matching,　this　paper　applies　a　semi-global　stereo　matching　method　to　match　corrected　left　and　right　perspective　images.　Because　there　are　noise　points　and　holes　in　the　matched　disparity　map,　which　affect　the　image　quality,　a　sample　block　filling　method　which　combines　mean　filtering　and　point-by-point　scanning　is　proposed　to　repair　the　image.　Then　a　gradient　priority　selection　mechanism　is　proposed　to　maintain　the　edge　structure　of　the　object　in　the　process　of　restoration.　Experimental　results　show　that　the　proposed　method　is　good　for　the　restoration　of　holes　and　noises　in　disparity　maps,　and　the　processing　speed　is　improved　by　about　30%　compared　with　the　traditional　Criminisi　algorithm.