A Ground Truth System for Radar Measurements of Humans

Radar simulations of human targets can be deployed to reduce the measurement effort linked to dataset generation for tasks such as gesture or activity classification. However, simulations require realistic human motion data in order to capture the dynamics of the simulated activities. For this purpose, this paper proposes a stereo camera-based system that enables simultaneous recording of radar data and the corresponding human pose ground truth. By introducing a camera-radar calibration procedure, the 3D human poses and the radar system are synchronized both in time and space. Thus, the system enables the one-by-one re-simulation of the captured measurements for the investigation of simulation quality or sensor studies. The performance of the calibration procedure and the feasibility of direct re-simulations is shown with measurements of an exemplary gesture. In addition, the straightforward extension of the proposed approach to radar sensor networks is demonstrated.
acceptedVersion