Target　recognition　is　a　fundamental　research　topic　in　the　process　of　robot　grasping.　In　this　paper,　we　proposed　an　algorithm　framework　for　object　recognition　based　on　local　naive　Bayes　nearest　neighbor　with　Kinect.　With　the　emergence　of　local　invariant　feature　detection,　the　method　based　on　local　invariant　features　gradually　becomes　the　mainstream.　Object　recognition　is　realized　through　the　feature　matching　of　the　model　in　the　current　scene　and　the　models　in　the　library　based　on　local　invariant　property.　Considering　the　number　of　models　in　the　library　may　be　as　many　as　dozens　or　even　hundreds,　I　divide　the　recognition　process　into　coarse　and　fine　recognition,　the　part　of　coarse　recognition　adopts　the　local　naive　Bayes　nearest　neighbor　algorithm,　just　search　for　a　number　of　the　nearest　neighbors　of　the　object　to　be　identified,　it　does　not　need　to　compare　all　models　in　the　model　library　one　by　one,　the　computational　complexity　of　the　model　increases　with　the　number　of　models　in　the　logarithmic　growth,　So　we　can　effectively　deal　with　the　situation　of　large　data　in　library.　The　process　of　fine　recognition　is　a　process　of　layers　of　verification,　it　mainly　includes　geometric　verification,　pose　verification,　projection　verification,　the　model　with　the　most　matching　points　will　be　used　as　the　final　recognition　result.　In　the　end,　a　variety　of　performances　were　tested　on　the　garage　willow　database　and　the　grasping　experiments　of　the　robot　arm　demonstrate　the　superiority　of　my　proposed　method.