A　method　is　presented　to　accurate　face-mating　point　cloud　registration　after　dealing　with　noise　point.　Point　cloud　registration　is　divided　into　two　parts,　firstly,　coarse　registration　is　applied　for　visual　point　cloud,　then　three　local　sufaces　of　overlap　cloud　region　are　selected　to　be　mating　calculated　after　denosing　base　on　least　squares　fitting,　at　last　accurate　splicing　parameters　of　translation　and　rotation　are　acquired　by　nonlinear　least　square.　This　algorithm　is　easy　to　deal　with　the　denosing,　has　faster　convergence　speed　and　higher　registration　accuracy.　Its　feasibility　is　proved　by　samples.　©　(2014)　Trans　Tech　Publications,　Switzerland.