To　enhance　the　application　of　deep　learning　networks　in　recognizing　multi-target　human　actions,　this　paper　prposes　an　innovative　multi-target　separation　algorithm　utilizing　Range-Doppler　Maps　(RDMs).　The　algorithm　translates　Frequency　Modulated　Continuous　Wave　(FMCW)　radar　data　into　visual　Ms　and　employs　a　threshold-based　method　to　distinguish　targets　across　successive　frames,　with　the　frame　count　optimized　according　to　the　modal　value.　By　projecting　the　M　onto　the　velocity　axis　and　aggregating　column-wise　data　incrementally,　a　micro-Doppler　Map　(mDM)　is　constructed.　Experiments　verify　the　algorithm＇s　effectiveness　with　minimal　impact　on　recognition　accuracy,　highlighted　by　an　average　Intersection　over　Union　(IoU)　score　of　92.53%.