Radar-Based Online Learning of Human Skeleton Keypoints Without Any Prior Real Data

Current techniques for radar-based human skeleton keypoint estimation require massive amounts of real data and exhibit large estimation errors because they only take into account positional error. This paper proposes a new method for skeleton estimation that eliminates the need for real training data prior to deployment by initially training models using only simulations, but fine-tuning model training online, after deployment in a room based on the new concept of kinematic cycle consistency (KCC). KCC involves defining a physics-based loss term from micro-Doppler signatures reconstructed from the estimated skeleton and essentially minimizes velocity-based kinematic errors. In this way, the model for skeleton estimation can be fine-tuned without needing any motion capture data for ground truth in an online fashion post-deployment. Our results show that our proposed approach results in improved skeleton estimation accuracy without using any prior real data in comparison to training with real data only. Accuracy is evaluated against ground truth established by a VICON optical tracking system.