Context-Aware Multimodal mmWave Radar and Vibration Dataset for Non-Intrusive and Privacy-Preserving Fall Detection in Bathrooms

This dataset contains tightly synchronized recordings from a full-scale bathroom mock-up designed to replicate residential conditions with wet, highly reflective surfaces and confined geometry.  Two privacy-preserving ambient sensors were used: (i) a millimeter-wave (mmWave) radar node that outputs frame-wise point clouds and kinematic summaries at 12.5 Hz;  and (ii) a floor-mounted triaxial vibration node sampled at 200 Hz (resampled to 100 Hz for modeling).Nine scenarios are included: empty bathroom, light object drop, heavy object drop, normal walking, bent-posture walking, wall-supported walking, quiet standing, squatting, and intentional falls.  Human activities were conducted with running water to emulate realistic wet environments.  Across all trials the corpus exceeds 3 hours, totaling \u22481.1\u00d710^5 radar frames and \u22483.1\u00d710^6 vibration samples.Each trial provides synchronized streams, frame-level annotations around impact (\u00b1250 ms tolerance), and subject-independent splits (60\/20\/20) for train\/validation\/test.  The dataset supports research on alignment-aware, temporally coherent multimodal fusion for fall detection, and accompanies our paper that introduces Motion\u2013Mamba (radar) and Impact\u2013Griffin (vibration) streams with low-rank bilinear coupling and Switch\u2013MoE fusion.