We propose to detect different human motions using planar phased-array FMCW radar through investigating 3D point clouds, which has not yet been presented. While human motion has been analyzed using micro-Doppler signatures, studies investigating an approach that employs point clouds are lacking. We measured 7 human motions including bowing, kicking, punching, walking, running, sitting down, and standing using a planar phased-array FMCW radar system operating at 77GHz. Next, 3D point clouds were extracted by calculating direction-of-arrival from point scatterers on the human body. As the point clouds contained human posture information, we classified the motions using convolutional neural networks. The classification accuracy was 80%.