RadProPoser: A Framework for Human Pose Estimation with Uncertainty Quantification from Raw Radar Data

Jonas Leo Mueller, Lukas Engel, Eva Dorschky, Daniel Krauss, Ingrid Ullmann, Martin Vossiek, Bjoern M. EskofierarXiv 2025 · DOI · arXiv

TL;DR

This paper introduces RadProPoser, a novel framework for estimating human 3D pose directly from raw radar data. It uniquely quantifies uncertainty in these estimations, making radar-based pose estimation more reliable and interpretable for safety-critical applications.

What Problem Does This Paper Address?

Human pose estimation (HPE) is crucial for many AI applications, but traditional methods often rely on visual data, which can be compromised by lighting conditions or privacy concerns. Radar offers a privacy-preserving, illumination-invariant alternative, but its data is inherently noisy and prone to multipath effects, making accurate HPE challenging. Existing radar-based HPE methods often struggle with quantifying the uncertainty of their predictions, which is vital for building trust and ensuring safety in high-risk domains. This paper aims to bridge this gap by developing an end-to-end system that not only estimates human pose from raw radar data but also provides reliable uncertainty estimates for each predicted joint.

Methods

RadProPoser employs a probabilistic encoder-decoder architecture. The core innovation lies in integrating variational inference into the keypoint regression process. This allows the model to jointly predict 26 3D joint locations and their associated heteroscedastic aleatoric uncertainties. The framework processes complex-valued radar tensors from a compact 3-transmitter, 4-receiver MIMO radar. It explores different probabilistic formulations for latent priors and likelihoods, including Gaussian and Laplace distributions. The model uses a shared encoder based on the HRNet architecture to process temporal radar inputs, followed by a self-attention mechanism to capture dependencies across time. A probabilistic decoder then outputs the pose and uncertainty estimates. For uncertainty quantification, they leverage a two-stage approach: first, modeling aleatoric uncertainty directly, and then recalibrating these predictions to estimate total uncertainty.

Datasets

The authors introduce a new dataset collected with optical motion capture (OMC) ground truth, synchronized with radar data. This dataset comprises recordings of 12 participants performing nine different actions at three distinct angles (0°, 45°, and 90°) relative to the radar. This new benchmark is crucial for evaluating radar-based HPE systems with accurate 3D ground truth.

Key Insights and Ideas

  • Uncertainty is Key for Radar HPE: The inherent noise and multipath effects in radar data make uncertainty quantification not just desirable but essential for reliable HPE. RadProPoser demonstrates that these uncertainties can be learned directly from the data.
  • End-to-End Processing of Raw Radar Data: By processing complex-valued radar tensors directly, the framework avoids information loss that can occur with aggressive preprocessing, leading to potentially more accurate and informative pose estimations.
  • Latent Space for Generative Tasks: The learned latent space representations are shown to be structured and informative, enabling effective data augmentation for downstream tasks like activity classification, boosting F1 scores significantly.
  • Angle Matters: Performance is notably better when the subject is at a 45° or 90° angle to the radar, suggesting that direct frontal views (0°) present unique challenges for radar-based pose estimation due to reduced visibility of certain body parts.

Key Results

RadProPoser achieves state-of-the-art performance on their newly released dataset, with an overall mean per-joint position error (MPJPE) of 6.425 cm, and an impressive 5.678 cm at the optimal 45° aspect angle. The learned uncertainties show strong alignment with actual pose errors and can be calibrated to produce reliable prediction intervals, with the best configuration achieving an expected calibration error (ECE) of 0.021. Furthermore, sampling from the latent distributions enabled data augmentation for activity classification, resulting in an F1 score of 0.870.

Limitations and Future Work

The authors acknowledge that their dataset was collected in a controlled laboratory environment, which may not fully capture real-world variability. They also note that their current setup focuses on single-person HPE and suggest future work on multi-person estimation. They plan to investigate the impact of different radar configurations, sensor fusion, and explore more advanced temporal modeling techniques like belief propagation. Additionally, they aim to use the learned uncertainties to guide the generation of synthetic radar data for more efficient training and to incorporate these uncertainties into musculoskeletal simulations for more robust motion reconstruction.

Summary updated

BibTeX
@misc{mueller2025radproposer,
  author = {Jonas Leo Mueller and Lukas Engel and Eva Dorschky and Daniel Krauss and Ingrid Ullmann and Martin Vossiek and Bjoern M. Eskofier},
  title = {{RadProPoser: A Framework for Human Pose Estimation with Uncertainty Quantification from Raw Radar Data}},
  year = {2025},
  eprint = {2508.03578},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2508.03578},
}
Figure