Due　to　image　noise,　illumination　and　occlusion,　to　get　an　accurate　and　dense　disparity　with　stereo　matching　is　still　a　challenge.　In　this　paper,　a　new　dense　stereo　matching　algorithm　is　proposed.　The　proposed　algorithm　first　use　cross-based　regions　to　compute　an　initial　disparity　map　which　can　deal　with　regions　with　less　or　similar　texture.　Secondly,　the　improved　hierarchical　belief　propagation　scheme　is　employed　to　optimize　the　initial　disparity　map.　Then　the　left-right　consistency　check　and　mean-shift　algorithm　are　used　to　handle　occlusions.　Finally,　a　local　high-confidence　strategy　is　used　to　refine　the　disparity　map.　Experiments　with　the　Middlebury　dataset　validate　the　proposed　algorithm.　©　2013　Springer-Verlag.