This study presents a novel real-time home health monitoring system based on milli-meter-wave FMCW radar, which provides a cost-effective and privacy-friendly solu-tion compared to RGB camera sensors. For the home health monitoring system, in addition to the four common human activities – walking, standing, sitting, and lying – it is crucial to promptly detect emergency situations. Therefore, we support fall detec-tion as well as the recognition of two gestures, one for headache and another for SOS, to notify emergency events. However, due to the sparsity of radar point clouds, we opt not to use raw radar point clouds directly as input for neural networks. Instead, we introduce a new efficient representation of motion waveforms by leveraging spa-tial and velocity time-series variation information from radar point clouds to repre-sent human activities, and train a lightweight 1D-CNN+LSTM neural network to achieve real-time recognition. These motion waveforms undergo enhancement through a pre-filtering process aiming at improving neural network performance. The experimental results show that our proposed method can achieve an overall recogni-tion accuracy of 94% at a throughput of 30 FPS on our self-collected 7492 data us-ing a lightweight laptop.