Synchronized raw radar and three-dimensional human pose data for prehabilitation exercises

We present a synchronised dataset of millimetre-wave (mmWave) radar measurements and camera-derived three-dimensional (3D) human pose annotations acquired during clinically relevant cancer prehabilitation exercises. Thirteen volunteers performed 19 standing activities targeting strength, flexibility, balance, and core stability, following a standardised recording protocol. Radar data were collected using a commercial off-the-shelf four-dimensional (4D) in-phase and quadrature (IQ) mmWave radar (Vayyar IMAGEVK-74) configured with 10 transmit antennas and 20 receivers, and operated in stepped-frequency continuous-wave (SFCW) mode over 62–66.5 GHz with 100 frequency tones, yielding a frame rate of approximately 13 Hz. In parallel, an RGB–D camera recorded video at 30 fps; 33-joint 3D skeletal landmarks were extracted offline using MediaPipe BlazePose Full, then remapped to a COCO-style 17-joint representation. Radar and pose streams were temporally aligned via nearest-neighbour timestamp matching. The provided 3D skeletal annotations represent relative joint positions expressed in a local, hip-centred coordinate system rather than absolute global positions. We provide subject-level files containing IQ radar data, camera timestamps with alignment indices, and 3D joint coordinates. We validate our recordings with an end-to-end deep-learning-based radar-to-skeleton model that achieves centimetre-level joint prediction performance on a held-out test set.