Large-scale　multiple-input　multiple-output　(MIMO)　system　with　tens　to　hundreds　of　antennas　is　getting　increasing　attention　for　its　very　high　spectral　efficiency　and　energy　efficiency.　However,　for　frequency　division　duplexing　(FDD)　system,　the　cost　of　CSIT　feedback　becomes　a　bottleneck　with　the　BS　antennas　increasing.　To　reduce　the　cost　of　the　downlink　training　and　the　CSIT　uplink　feedback,　joint　spatial　division　and　multiplexing　(JSDM)　scheme　is　proposed　by　Ansuman　Adhikary,　et　al.,　which　divides　users　into　different　groups　according　to　their　channel　covariance　eigenvectors　and　adopts　the　two-stage　precoding.　In　this　paper,　a　Density　Based　Spatial　Clustering　of　Application　with　Noise　(DBSCAN)　user　grouping　algorithm　is　proposed　for　JSDM　scheme,　which　divides　all　the　users　based　on　the　density　distribution　of　users.　DBSCAN　can　find　the　groups　with　arbitrary　shape　and　size,　while　DBSCAN　has　two　parameters　which　is　hard　to　determine.　In　this　paper　the　parameters　are　determined　by　detecting　the　density　distribution　using　k-dist　graph.　The　simulation　results　show　that　the　proposed　DBSCAN　grouping　algorithm　has　better　performance　than　K-Means　and　K-Medoids.