As　we　are　getting　deeper　into　the　world　of　human-computer　interaction,　the　human　activity　recognition　technology　becomes　more　and　more　important.　Due　to　the　complexity　of　the　actual　environment,　we　need　a　more　reliable　and　powerful　system　that　can　recognize　human　activities　in　a　variety　of　environments　with　high　accuracy.　Taking　these　factors　into　consideration,　we　propose　a　human　activity　recognition　method　based　on　data　fusion　of　the　FMCW　radar　and　image,　which　uses　the　complementarity　of　the　different　data　to　improve　the　performance　of　the　system.　Apart　from　this,　the　domain　adaptation　is　applied　to　reduce　the　data　differences　caused　by　the　changes　of　environments　and　user　habits　in　the　practical　application.　Finally,　we　have　implemented　real-time　applications　based　on　the　proposed　algorithm　on　the　edge　computing　platform.　The　experimental　results　show　that　the　recognition　accuracy　of　the　fusion　system　can　reach　98.7%,　and　the　average　running　time　of　the　real-time　system　is　about　0.17s.　©　2021　IEEE.