Human Activity Recognition (HAR) is crucial in areas such as healthcare and intelligent security, particularly for elderly care and remote monitoring. Traditional methods using cameras or wearables face issues like privacy concerns and inconsistent usage. To overcome these challenges, we propose mmPoint-GBG, a framework that leverages millimeter-wave radar to capture point cloud data of human activities. This data is directly processed using a dynamic spatiotemporal Graph Neural Network (GNN), which integrates temporal features with a Bidirectional Gated Recurrent Unit (Bi-GRU) and performs classification through fully connected layers. On the MMActivity dataset, mmPoint-GBG achieved a 97.02% recognition accuracy while enhancing privacy protection, environmental adaptability, and computational efficiency.