---
title: 'RadProPoser: Probabilistic Radar Pose Estimation'
url: https://www.emergentmind.com/topics/radproposer
type: topic
---

# RadProPoser: Probabilistic Radar Pose Estimation

RadProPoser is a probabilistic encoder–decoder framework for single-person, three-dimensional human pose estimation from raw millimeter-wave radar that jointly regresses 26 joint locations and per-joint uncertainty from minimally processed complex-valued radar tensors [2508.03578]. It was introduced to address a central limitation of radar-based human pose estimation: although radar is privacy-preserving and illumination-invariant, its measurements are strongly affected by thermal, quantization, and phase noise, as well as multipath propagation, refraction, and scattering. RadProPoser therefore treats pose estimation as a probabilistic regression problem rather than a purely deterministic coordinate prediction task. Its reported results include an overall mean per-joint position error of 6.425 cm, 5.678 cm at the 45° aspect angle, and a best calibrated expected calibration error of 0.021, together with a downstream activity-classification result of \(F1 = 0.870\) when latent sampling is used for data augmentation [2508.03578].

## 1. Problem formulation and sensing regime

RadProPoser addresses single-person 3D human pose estimation from raw radar tensors. The target output is a 26-keypoint skeleton in 3D Euclidean coordinates, supervised with optical motion-capture ground truth in meters and expressed in the OMC world frame aligned to the recording space. For each joint \(j\) and axis \(k \in \{x,y,z\}\), the model predicts a mean \(\mu_{jk}\) and an uncertainty parameter, either \(\sigma_{jk}\) for Gaussian likelihoods or \(b_{jk}\) for Laplace likelihoods. A diagonal independence assumption or a full-covariance multivariate Gaussian can be used, and no explicit kinematic priors are imposed [2508.03578].

The sensing motivation is explicitly tied to radar’s operational advantages and liabilities. Radar-based HPE enables non-contact motion sensing that is robust to lighting and more privacy-preserving than RGB, but the measurements are degraded by environment-dependent artifacts. RadProPoser is designed around the observation that these artifacts produce overconfident point estimates if training relies only on point losses. The framework therefore models heteroscedastic aleatoric uncertainty per joint and axis and later recalibrates predictive distributions to total uncertainty.

A frequent misunderstanding is to regard radar uncertainty estimates as merely ancillary diagnostics. In RadProPoser, uncertainty is part of the primary prediction target: the system is intended not only to estimate pose, but also to characterize when and where the estimate is unreliable. This is central to the paper’s claim that explainable and reliable human motion analysis in radar applications requires calibrated probabilistic outputs rather than only low MPJPE.

## 2. Radar acquisition, tensor representation, and preprocessing

The sensor is a Texas Instruments IWR6843AOPEVM FMCW radar at 60 GHz with three transmit and four receive antennas. Time-division multiplexing activates transmitters sequentially, yielding 12 TDM-MIMO virtual channels in an L-shaped virtual antenna array in azimuth-elevation. The hardware configuration uses RF bandwidth of approximately 1.02 GHz, a frame rate of 15 Hz, chirp duration of approximately \(17\,\mu s\), ADC sampling frequency of 3.8 MHz, 64 samples per chirp, and 128 chirps per frame [2508.03578].

RadProPoser ingests FFT-processed radar cubes that preserve complex values and are described as “raw” in the sense of minimal preprocessing. Static clutter is attenuated by centering in the chirp dimension prior to the FFT. A 4D FFT is then applied across ADC sample, chirp, azimuth, and elevation to obtain range, Doppler, azimuth, and elevation bins. The representation remains in polar coordinates rather than being converted to Cartesian.

The tensorization is explicit. The original complex-valued radar cube is
$$
X \in \mathbb{C}^{2 \times 8 \times 4 \times 4 \times 64 \times 128},
$$
with dimensions corresponding to \([{\rm real/imag}, T=8, {\rm azimuth}, {\rm elevation}, {\rm ADC}, {\rm chirps}]\). After FFT and reordering, the tensor becomes
$$
I \in \mathbb{C}^{2 \times 8 \times 128 \times 4 \times 4 \times 64},
$$
arranged as \([{\rm real/imag}, {\rm time}, {\rm Doppler}, {\rm azimuth}, {\rm elevation}, {\rm range}]\). Real and imaginary components are then concatenated along the time dimension to produce
$$
F \in \mathbb{R}^{16 \times 128 \times 4 \times 4 \times 64}.
$$

This representation is methodologically important because many prior radar HPE pipelines compress radar to point clouds via CFAR or other intermediate abstractions. RadProPoser instead keeps minimally processed tensor structure available to the network, which allows uncertainty modeling to operate on the same raw measurement domain as pose inference.

## 3. Architecture and latent-variable design

The encoder is a shared 3D HRNet backbone adapted from prior HRNet-based radar pose pipelines. Each temporal slice \(F^t \in \mathbb{R}^{1 \times 128 \times 4 \times 4 \times 64}\) is processed independently through the shared backbone. HRNet maintains high-resolution streams via parallel branches with strided convolutions, and multi-resolution features are upsampled and concatenated to avoid information loss in the final stage. Two additional 3D convolution layers with batch normalization and ReLU reduce the channels to \(F^t \in \mathbb{R}^{1 \times 8 \times 4 \times 4 \times 1}\) [2508.03578].

Temporal fusion is performed after flattening spatial dimensions to obtain \(F \in \mathbb{R}^{16 \times 128}\). Self-attention is then applied over time using
$$
Q = F W_Q,\qquad K = F W_K,\qquad V = F W_V,
$$
$$
{\rm Attention}(Q,K,V) = {\rm softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)V,
$$
followed by a residual connection \(F' = F + {\rm Attention}(Q,K,V)\) and a linear layer. The latent dimension is \(d_{\rm lat} = 256\), which was the best-performing setting in the ablation study.

The decoder is a lightweight two-layer MLP that maps latent samples to pose-distribution parameters. Sampling is vectorized by stacking samples along the batch dimension:
$$
z \in \mathbb{R}^{500 \times 256}, \qquad \hat{y} \in \mathbb{R}^{500 \times 78},
$$
with the 78 outputs corresponding to \(26 \times 3\) coordinates or associated distribution parameters. A full covariance \(\Sigma \in \mathbb{R}^{78 \times 78}\) can also be estimated.

The architectural choice is not only a matter of accuracy. It encodes a division of labor: HRNet retains fine-grained radar structure, temporal attention aggregates short motion context, and the variational decoder converts this representation into a distribution over body poses rather than a single coordinate vector. This suggests that RadProPoser’s main innovation lies in integrating uncertainty quantification into the representational core of raw-radar pose estimation, not merely appending a calibration layer to an otherwise standard regressor.

## 4. Probabilistic inference, heteroscedastic losses, and calibration

RadProPoser explores both Gaussian and Laplace latent priors and likelihoods. The Gaussian latent posterior is
$$
q_\phi(z|x) = \mathcal{N}(\mu_l, {\rm diag}(\sigma_l^2)),
$$
with reparameterization
$$
z = \mu_l + \alpha(\sigma_l \odot \epsilon), \qquad \epsilon \sim \mathcal{N}(0,I),
$$
where \(\alpha\) is a learnable global scale. The Laplace alternative is
$$
q_\phi(z|x) = {\rm Laplace}(\mu_l,b_l),
$$
with \(b_l\) made positive by softplus and sampled through inverse-CDF reparameterization [2508.03578].

The variational objective is written as
$$
L_{\rm ELBO} = \mathbb{E}_{q_\phi(z|x)}[\log p_\theta(y|z,x)] - KL(q_\phi(z|x)\|p(z)),
$$
and implemented in practice as \(L = {\rm NLL} + \beta\,KL\). For Gaussian latent priors, a closed-form KL to \(\mathcal{N}(0,I)\) is used. For Laplace latent priors, the paper reports training via NLL without a closed-form Laplace KL term.

The heteroscedastic Gaussian negative log-likelihood is
$$
L_{\rm Gauss} = \sum_{j,k} \left[\frac{\|y_{jk}-\mu_{jk}\|^2}{2\sigma_{jk}^2} + \frac{1}{2}\ln \sigma_{jk}^2\right],
$$
and the implementation uses a \(\gamma\)-weighted log-variance term for stabilization. The Laplace counterpart is
$$
L_{\rm Laplace} = \sum_{j,k} \left[\frac{\|y_{jk}-\mu_{jk}\|}{b_{jk}} + \ln(2b_{jk})\right].
$$
For the full-covariance Gaussian output, the multivariate NLL is
$$
{\rm NLL}(y,\mu,\Sigma) = \gamma \ln |\Sigma| + (y-\mu)^T \Sigma^{-1}(y-\mu).
$$

Uncertainty treatment proceeds in two stages. First, the model predicts heteroscedastic aleatoric uncertainty, with per-joint uncertainty
$$
u_k = \sum_{d=1}^{3}\sigma_{k,d}^2.
$$
Second, the paper follows Kuleshov et al. and uses isotonic regression on one held-out participant to learn a mapping \(R:[0,1]\rightarrow[0,1]\), which is then applied to the other two test participants to convert aleatoric outputs into calibrated total uncertainty. This explicitly separates learned data-dependent noise from post-hoc total predictive uncertainty. The conceptual decomposition is given by the law of total variance,
$$
{\rm Var}(y|x) = \mathbb{E}_w[{\rm Var}(y|x,w)] + {\rm Var}_w[\mathbb{E}(y|x,w)].
$$

A common misconception is that the most accurate probabilistic variant must also be the best calibrated. RadProPoser demonstrates the opposite: the Gaussian latent with full-covariance Gaussian likelihood gives the lowest MPJPE, whereas the Gaussian latent with Laplace likelihood gives the best calibrated ECE. Accuracy and calibration are therefore treated as distinct empirical properties rather than interchangeable ones.

## 5. Dataset, training protocol, and empirical results

The dataset contains 12 participants, of whom 4 are female and 8 male, with age \(28.92 \pm 3.15\) years, height \(176.25 \pm 6.25\) cm, and weight \(72.33 \pm 12.37\) kg. It includes nine exercises: left, right, and bilateral upper-limb extension, bicep curls, front arm rotation, trunk forward bending, left and right front lunge, and squats. Data were recorded at view angles of \(0^\circ\), \(45^\circ\), and \(90^\circ\), with each sequence lasting 36 seconds and repeated three times per angle. Ground truth comes from an OptiTrack Flex 13 OMC system with 12 cameras and Motive:Body software, and radar and OMC are synchronized at 15 Hz with a hardware trigger [2508.03578].

The participant-wise split uses three held out for testing, one for hyperparameter tuning, and eight for training; after tuning, the model is retrained on nine participants. The best ablation setting uses an 8-frame input sequence, approximately 0.5 seconds, and latent dimension 256. Training uses 500 latent draws per example. Softplus is used for Laplace scales, \(\beta\)-weighted KL for Gaussian latents, \(\gamma\)-weighted log-variance regularization in the NLL, and an optional full-covariance Gaussian likelihood.

| Configuration or task | Metric | Result |
|---|---|---|
| Gaussian latent + Gaussian likelihood with full covariance | Overall MPJPE | 6.425 cm |
| Gaussian latent + Gaussian likelihood with full covariance | MPJPE at \(45^\circ\) | 5.678 cm |
| Gaussian latent + Laplace likelihood | Best calibrated ECE | 0.021 |
| Latent-sampled downstream activity classification | \(F1\) | 0.870 |

MPJPE is defined as
$$
{\rm MPJPE}=\frac{1}{NJ}\sum_{n=1}^{N}\sum_{j=1}^{J}\|p^{\rm pred}_{n,j}-p^{\rm gt}_{n,j}\|_2.
$$
Uncalibrated ECE values range across variants, with Normalizing Flows giving the best uncalibrated ECE of 0.078 and Gaussian–Gaussian–Cov yielding 0.140, but isotonic recalibration improves all variants. Per-keypoint analysis shows higher errors and uncertainties for smaller reflective surfaces, specifically hands, hand-ends, forearms, and head, and a systematic left–right asymmetry attributed to viewing direction in the study design.

The framework also supports downstream activity classification through latent-space augmentation. Entire exercise sequences are encoded to latent sequences, 100 samples per sequence are generated for each participant using 11 \(\alpha\) values, and a 1D CNN classifier compresses \(256 \rightarrow 64\) channels, average-pools over time, and outputs logits through a fully connected layer. The resulting augmented model achieves precision 0.891, recall 0.852, accuracy 0.852, and \(F1 = 0.870\), compared with a non-augmented baseline \(F1 = 0.657\), precision 0.688, recall 0.627, and accuracy 0.630.

## 6. Comparative position, limitations, and significance

RadProPoser is positioned against prior radar HPE systems that either compress radar to point clouds or process full radar cubes without explicit per-joint uncertainty regression. The paper further contrasts it with two-radar pipelines that detect 2D keypoints and lift to 3D, and with a 12×16 array tensor-based system reporting MPJPE of approximately 9.91 cm across diverse scenarios. RadProPoser reports 6.425 cm using only 3 transmitters and 4 receivers, while explicitly modeling and calibrating uncertainty [2508.03578].

Its stated novelty is precise: it is presented as the first end-to-end raw radar tensor HPE system to explicitly regress per-joint heteroscedastic uncertainty and calibrate it to total uncertainty, with dataset release and downstream generative utility. The paper’s broader significance lies in showing that uncertainty can be actionable rather than decorative: it identifies difficult joints, produces reliability diagrams and prediction intervals, and supports data augmentation for downstream recognition.

The limitations are equally explicit. The study is restricted to a single person in a controlled laboratory. Front views at \(0^\circ\) show higher variability. Joints with weak radar reflections, especially hands and head, have higher uncertainties. Runtime and model size are not explicitly reported. Deployment is described as near-real-time on modern GPUs, but calibration is recommended to be recomputed per environment or augmented with conformal methods for field use.

Several misconceptions are therefore addressed by the evidence. Privacy preservation does not remove viewpoint dependence. Raw radar tensors do not eliminate the need for temporal context, as the best results use 8 frames rather than a single frame. Most importantly, uncertainty quantification does not by itself guarantee trustworthy predictions; the paper requires post-hoc isotonic recalibration to obtain calibrated total uncertainty. This suggests that RadProPoser should be understood not as a deterministic pose regressor with auxiliary confidence scores, but as a probabilistic radar sensing framework in which prediction, calibration, and downstream decision support are inseparable components.

Source: https://www.emergentmind.com/topics/radproposer