This paper tackles the challenge of achieving high-accuracy classification of human trajectories in complex indoor environments using millimeter-wave radar point clouds, while preserving privacy. After enhancing raw point-cloud motion signals via median and MTI filtering, we extract multi-dimensional physical features—including position, velocity, and radar reflection intensity—to form robust trajectory descriptors. An XGBoost-based classifier is then employed to discriminate between different motion patterns with over 96% overall accuracy. Finally, a Perceiver IO model captures spatio-temporal dependencies of classified trajectories to predict short-term target positions. Experiments demonstrate that the proposed system delivers both precise classification and sub-5 cm prediction error, validating its effectiveness and robustness for real-time indoor monitoring.