Skinned Scatterers: Predicting Human-Induced Fading from Body Motion

Wireless engineers usually treat the people walking through a radio link as noise. In the standard fading models, a moving body either adds random diffuse scattering or blocks the link as a solid cylinder. The result is a channel whose statistics are fixed in advance: a Rician channel, meaning one strong steady path plus random scatter, with a constant ratio between the two (the $K$-factor) and a textbook Doppler spread. Human motion does not behave like that. The arm and leg swings of walking produce structured, repeatable Doppler signatures, and WiFi alone carries enough of that structure to recover a 17-joint skeleton to within about 3 cm (MDPose). This proposal turns that observation around. If body motion is legible in the channel, then a model of body motion should be able to predict the channel, and with it the fading statistics that drive outage, retransmission and link adaptation.

The idea in two sentences: represent a moving person as a cloud of small radio scatterers attached to a skeleton the way skin is attached to bones in character animation, add their geometrically phased echoes to a learned model of the static room, and read off both the instantaneous channel and its time-varying fading statistics. Because the scatterers deform smoothly with the body instead of moving as rigid blocks, the same fitted model should transfer to new activities, new people and new antenna placements. That transfer is where black-box predictors and rigid-cylinder blockage models are expected to fail.

Core Claim

The claim is that a forward channel model driven by body pose and built on non-rigid scatterers yields fading statistics that generalize across motions in a way the two existing families do not. The first family is the stationary fading models used in network analysis. The second is the static or transceiver-only-deforming wireless radiance fields, which are learned 3D maps of how radio energy propagates through a room. Concretely, the received channel at frequency $f$ and time $t$ is modeled as

$$ h(f,t) = h_{\mathrm{env}}(f) + \sum_{i=1}^{N} a_i \, \exp\!\Big(-j\,2\pi f\,\tfrac{\lVert x_i(t)-p_{\mathrm{tx}}\rVert + \lVert x_i(t)-p_{\mathrm{rx}}\rVert}{c}\Big), $$

Here $h_{\mathrm{env}}$ is the static-room channel, taken from a pretrained wireless radiance field. Each of the $N$ body scatterers has a learned complex reflectivity $a_i$ and a position $x_i(t)$, obtained by blending the transforms of nearby skeleton bones with learned weights (linear blend skinning). $p_{\mathrm{tx}}, p_{\mathrm{rx}}$ are the antenna positions and $c$ is the speed of light. A shadowing term, also driven by the skinned geometry, attenuates $h_{\mathrm{env}}$ when the body occludes the direct path.

Micro-Doppler is then not learned separately. It is the time derivative of the path lengths above, so the Doppler spectrogram is a prediction and not an input. The fading description follows by windowing. Over a window $W$, the instantaneous $K$-factor is

$$ K_W(t)=\frac{\lvert \bar h_W(t)\rvert^2}{\frac{1}{|W|}\sum_{\tau\in W}\lvert h(\tau)-\bar h_W(t)\rvert^2}, $$

where $\bar h_W(t)$ is the window mean. Outage probabilities, level-crossing rates and fade durations are computed from samples of $h$ under a distribution of body motions, for example motion-capture sequences of a given activity. The modeling contribution is a mapping from motion to fading statistics; no such mapping exists in the file.

Relation to Prior Work

Three wireless radiance field papers supply the static-room machinery and also mark the gap. NeWRF learns amplitude and phase along rays and treats wall reflections as mirror-image "virtual transmitters", but it assumes a frozen room of flat surfaces. NeRF-APT swaps the backbone for an attention U-Net and names dynamic obstacles as an open limitation. SwiftWRF renders the receiver's angular spectrum with 2D Gaussians and adds a deformation network, but that network is driven only by where the transceiver is. We reuse one of these as $h_{\mathrm{env}}$ and add what they lack: scatterers that deform with an external articulated body.

MDPose motivates the work by showing that human micro-Doppler is structured. It solves the inverse problem, going from RF signals to pose, and says nothing about fading experienced by other links. We solve the forward problem and use MDPose's estimator as a consistency probe.

On the vision side, Superpoint Gaussian Splatting and RigidFlow constrain the design. They show that grouping a moving surface into rigid clusters creates seams at cluster boundaries and fails under stretching. Are NeRFs ready for autonomous driving? reports the same failure for humans approximated as rigid boxes in sensor simulation. These results justify skinning rather than per-limb rigid primitives, but they are visual evidence. Whether the seams matter at radio wavelengths is exactly what our ablation tests.

AlAmmouri et al. show that collapsing structured specular components into diffuse Gaussian fading distorts interference and coverage analysis. That result motivates replacing a "depressed-$K$ Rician plus classic Doppler spread" description of people with a motion-derived one.

Research Program

The build has three parts.

First, fit a static-room field ($h_{\mathrm{env}}$) using the SwiftWRF or NeRF-APT recipe on empty-room measurements.

Second, attach a few hundred to a few thousand scatterers to a skeleton. Their bone-blending weights and complex reflectivities are learned from paired channel and pose recordings. The weights are initialized from body-surface proximity and regularized for smoothness across neighbouring scatterers, which prevents the model from collapsing into rigid per-bone clusters.

Third, add a differentiable occlusion term along the direct and dominant reflected paths, computed from the same skinned geometry.

The analysis step drives the fitted model with motion-capture sequences from Human3.6M and CMU MOCAP to produce families of $K_W(t)$ curves, Doppler spectra and fade-duration distributions per activity and per antenna geometry. We then ask whether these collapse onto a low-dimensional parametric fading family. For example, we test whether a Rician model whose $K$ and Doppler parameters are periodic in the gait cycle suffices. If it does, that compact family is itself a result usable in stochastic-geometry analysis. If it does not, the deviation tells analysts what the classical models miss.

Data is the honest bottleneck. The file lists no benchmark with paired channel matrices, ground-truth body pose and fading labels. The program therefore uses two sources. The first is a synthetic stage in which ground truth comes from physical-optics scattering off a dense, smoothly deforming body mesh driven by motion capture, at 5.8 GHz and at a millimetre-wave band. The second is a real stage on the MDPose passive WiFi radar dataset, which has 7 subjects, 3 environments, about 190 minutes of recordings and Kinect skeletons. There, only Doppler spectrograms and channel magnitude are checkable.

SimHumalator, the mocap-to-micro-Doppler simulator used by MDPose, builds bodies from primitives. It is therefore treated as a baseline and never as ground truth.

Baselines

The naive reference is the classical stationary description: a Rician channel whose $K$-factor and Doppler spread are fitted once per environment from the training recordings, ignoring pose. Any proposal in this area must beat it on fade-duration and outage-curve error. If it does not, the whole motivation fails.

A stronger statistical baseline fits $K$ and Doppler spread as smooth functions of time per training sequence, then predicts test sequences by matching activity labels. This tests whether "activity-aware statistics" without geometry is enough.

The strongest existing method is SwiftWRF, with its deformation network extended to take the skeleton joint positions as additional input alongside the transceiver coordinates. This is the most direct competitor: it already has fast rendering, learned deformation and complex-valued attributes, and conditioning it on pose is the obvious one-line extension. Its reported static-scene results are a median RSSI L1 error of 2.91 dB on NeRF2-s23 and a PSNR gain of 2.33–10.41 dB over earlier baselines. For context, NeRF-APT reports 24.91 dB CSI SNR on the ArgosV2 64-antenna dataset. There is no reported result found on Lacuna for any method on human-induced fading statistics or pose-conditioned channel prediction. All numbers in the dynamic setting will therefore be produced by us, with every method given the same training pairs, the same static-room initialization, equal parameter count within a factor of two, and the same hyperparameter search budget.

A rigid-primitive baseline represents each limb as a rigid capsule carrying the same learnable scatterers, moving with its bone. It is trained identically.

The key ablation removes skinning but keeps everything else. The same scatterers become hard-assigned to their nearest bone, which isolates whether non-rigid blending, as opposed to merely having pose-driven geometry, is what helps.

Beating the strongest baseline means two things. On held-out activities, held-out subjects and held-out antenna placements, our model must have lower spectrogram error and lower divergence between predicted and true fade-duration distributions. It must also not lose on the in-distribution split.

Why not just…?

Why not just condition SwiftWRF's deformation network on pose? This is the real threat. On in-distribution splits it may match us, since a flexible network can memorize how familiar poses map to channels. The bet is about transfer. Because our scatterer phases are tied to explicit path lengths, a new antenna placement or an unseen activity changes the prediction in a physically determined way, while the black-box network must extrapolate. NeRF-APT's finding that unstructured MLP and VAE baselines collapse (spectrum SSIM 0.11 against 0.866; RSSI RMSE 11.26 dB against 3.51 dB) suggests that structure matters for radio fields, but it says nothing directly about dynamic scenes. The held-out-placement and held-out-activity splits settle the question. If the pose-conditioned SwiftWRF matches us there too, the paper's claim reduces to the statistical analysis and we would report that.

Why not just use rigid cylinders, as ray tracers do? At 5.8 GHz the wavelength is about 5 cm, comparable to the seam errors rigid limbs produce, so rigid primitives may well suffice at WiFi frequencies. We expect the gap to open at millimetre-wave, where elbow and knee folds change shadow edges by many wavelengths. The skinning ablation is run at both bands. If rigid primitives tie us at both, the non-rigid ingredient is unnecessary and the contribution narrows to the forward motion-to-fading mapping.

Why not just fit a Rician model with a lowered $K$-factor for crowded rooms? The cheapest alternative assumes body echoes add up to Gaussian clutter. MDPose's recovery of joint trajectories from WiFi is direct evidence that the echoes are structured rather than Gaussian. The naive baseline quantifies how much this matters for outage and fade duration, which are what link adaptation actually uses.

Proposed Experiments and Decisive Checks

On the synthetic stage, we train on a subset of Human3.6M subjects and activities and test on held-out ones. We then retest with antennas moved to positions unseen during training. Metrics are channel SNR in dB (as in NeRF-APT), Doppler spectrogram error, error in the $K_W(t)$ trajectory, and divergence between predicted and true fade-duration and level-crossing distributions, all at both frequency bands.

On the MDPose dataset, we train on two environments and test on the third, and hold out subjects. Here the metrics are spectrogram fidelity and channel-magnitude error.

The decisive cross-check feeds our synthesized spectrograms for held-out motions into an MDPose-style pose estimator trained on real data. If the forward model is right, the recovered skeleton error should approach the 29.4 mm MDPose reports on real signals, and it should degrade markedly for the rigid-primitive and pose-conditioned SwiftWRF syntheses. The second decisive check is the outage curve. For a fixed link budget, the naive stationary Rician model, the rigid baseline and our model each predict an outage probability under held-out walking motion, and we compare those predictions against the ground-truth channel samples. A clear win in outage error on held-out placements, not just in channel reconstruction, is the result that would justify adopting the approach.