According　to　the　requirement　of　low　power　consumption　and　high　accuracy　of　fall　detection　systems　for　the　elderly　in　a　community,　a　human　body　activity　model　based　on　tri-axial　acceleration　and　angular　velocity　is　first　established.　A　sensor　board　integrated　with　low-power　ZigBee　and　the　MPU6050　sensor,　which　can　sample　and　cache　data　in　sleep　mode,　is　developed.　An　interrupt-driven　algorithm,　which　can　collect　and　transmit　the　tri-axial　acceleration　and　the　angular　velocity　data　of　human　activities　through　ZigBee　technology　under　low　power,　is　designed.　Second,　the　data　with　regard　to　the　tri-axial　acceleration　and　angular　velocity　of　human　activities　are　received　in　real　time　by　a　server　via　the　sliding　window,　and　the　range　specification　of　the　data　is　correspondingly　mapped　into　three　channel　red-green-blue(RGB)　pixels　to　realize　the　image　representation　of　the　data.　Finally,　on　the　basis　of　analyzing　and　comparing　the　differences　between　activities　of　daily　living(ADLs)　and　fall　images,　a　fall　detection　convolutional　neural　network(FD-CNN)algorithm　is　designed;　the　network　is　trained　using　ADLs　and　fall　data　published　on　the　Internet.　The　experimental　results　show　that　the　accuracy　of　the　FD-CNN　algorithm　is　98.6%,　and　its　sensitivity　and　specificity　are　98.6%　and　99.8%,　respectively.　Compared　with　the　existing　fall　detection　algorithms,　the　FD-CNN　algorithm　has　significantly　superior　accuracy,　sensitivity,　and　specificity　in　terms　of　fall　detection.　Meanwhile,　the　system　presented　in　this　study　has　significant　features,　such　as　long　transmission　distance,　easy　networking,　and　lower　power　consumption,　in　comparison　to　the　Bluetooth-based　fall　detection　system.　Hence,　the　proposed　system　is　very　suitable　for　fall　detection　and　alarm　applications　for　the　elderly　in　a　community.　©　2019,　Editorial　Board　of　Journal　of　Tianjin　University(Science　and　Technology).　All　right　reserved.