According　to　the　requirement　of　low　power　and　accuracy　for　fall　detection,　an　activity　model　based　on　three-dimensional　attitude　angles　is　introduced,　and　the　difference　in　attitude　angles　and　signal　vector　magnitude　of　acceleration　between　daily　activities　and　falls　are　compared.　Second,　a　sensor　board　integrated　with　MPU6050　and　ZigBee　which　can　collect　and　transmit　the　tri-axial　accelerations　and　angular　velocities　of　human　activities　to　the　server　at　low　-power　is　developed.　Finally,　a　fall　detection　system　miming　on　the　server　is　developed　via　a　Kalman　filter　and　kNN　algorithm.　It　is　proved　by　experiment　that　the　accuracy　of　the　system　is　98.2%,　while　its　sensitivity　and　specificity　are　96.2%,　and　99.2%,　respectively.