RadPose: ISAC-based RF 3D Human Pose Estimation

RadPose: ISAC-based RF 3D Human Pose Estimation Geo Jie Zhou Kunzhe Song Ruxin Lin Huacheng Zeng Department of Computer Science and Engineering, Michigan State University

arXiv:2611.01400v1 [cs.CV] 1 Nov 2026

{geozhou, songkunz, linruxin, hzeng}@msu.edu

Abstract Human pose estimation under adverse lighting and occlusion conditions remains challenging for camera- and LiDAR-based systems. Although mmWave radars provide robustness to illumination and occlusion, they rely on dedicated hardware and licensed spectrum, limiting scalability. In this paper, we present RadPose, a learning-based monostatic integrated sensing and communication (ISAC) framework that enables mmWave communication devices to perform 3D human pose estimation. Unlike specialized radars, RadPose leverages monostatic full-duplex operation to extract channel impulse responses (CIRs) directly from standard OFDM communication signals, generating 3D radio point clouds without additional sensing hardware. To address the sparsity and noise of RF observations, RadPose employs a cross-frame transformer-based pose inference network with semantic-guided spatial attention and targetdriven temporal attention for robust spatio-temporal aggregation. It further adopts a joint-aware inverse kinematics refinement module based on graph convolution to enforce skeletal topology, bone-length consistency, and anatomical plausibility. Extensive experiments demonstrate RadPose achieves competitive accuracy compared to state-of-theart radar-based approaches while maintaining robust performance in unseen and occluded scenarios. The dataset and source code are available at https://github.com/GeoChou/RadPose.

1. Introduction Human pose estimation is a fundamental problem in computer vision [78]. Accurate 3D pose estimation allows machines to interpret human actions, infer intent, and interact safely in dynamic environments. Existing solutions predominantly rely on vision sensors such as cameras and LiDAR, which have driven remarkable progress in recent years [24, 44, 50]. However, both camera and LiDAR sensors depend on light propagation and reflection, making them vulnerable to poor illumination, adverse weather, and severe occlusions in cluttered scenes [8, 14, 29, 44]. Millimeter-wave (mmWave) radar has emerged as a

Table 1. Comparison of RadPose and other sensing modalities. Tech.

Medium

Waveform

Occlusion Resilience

Spectrum License

Hardware

Power Consump.

Cost

# of Devices Needed

Camera LiDAR Radar

Light Laser Radio WiFi CSI RadPose

— Pulsed FMCW OFDM OFDM

Poor Poor Good Good Good

— — Required Not required Not required

Dedicated Dedicated Dedicated Reuse Reuse

Low High Med ∼Zero ∼Zero

Low High Med ∼Zero ∼Zero

1 1 1 ≥2 1

ISAC

compelling alternative for human pose estimation, with prior work demonstrating remarkable performance [10, 23, 65]. Unlike camera and LiDAR, radar operates using radiofrequency (RF) waves, enabling robustness to lighting conditions and partial penetration through visual occlusions such as clothing, smoke, or lightweight obstacles. This property makes radar attractive for all-weather perception. Nevertheless, conventional radar systems typically require multi-GHz spectrum bandwidth to achieve the desired ranging resolution for pose reconstruction. Such bandwidth either demands dedicated licensed spectrum from regulatory bodies like the Federal Communications Commission (FCC) or must share spectrum with incumbent wireless systems, raising concerns about mutual interference with WiFi or cellular networks. Additionally, deploying standalone radar hardware increases system cost and limits scalability for widespread adoption [28]. In this paper, we introduce RadPose, a learning-based integrated sensing and communication (ISAC) framework that enables individual mmWave communication devices to estimate human pose within their field of view. Table 1 provides a comparison between RadPose and other sensing modalities. By embedding sensing capabilities directly into communication nodes, RadPose eliminates the need for dedicated radar spectrum licenses, specialized sensing hardware, or separate radio emissions solely for sensing purposes. Instead, it reuses OFDM communication signals for environmental perception, avoiding harmful interference to incumbent systems while enabling ubiquitous, low-cost deployment. This integration transforms widelydeployed wireless communication infrastructure into a distributed sensing network, opening a pathway toward scalable human pose estimation in scenarios where vision sensors fail or privacy is a constraint. RadPose comprises three key components. The first component is a reliable sensing data acquisition mecha-

nism. Most existing ISAC-based sensing systems adopt a bistatic configuration, where physically separated transmitter and receiver nodes collect channel state information (CSI). In such settings, dynamic variations in the transmitter-receiver distance, pose, and orientation introduce uncertainties in the measured CSI, limiting generalization across environments. More importantly, bistatic CSI measurements lack ranging information, making object distance estimation impossible. To address these challenges, RadPose enables monostatic full-duplex operation on mmWave communication devices, leveraging the strong isolation between transmit and receive antenna arrays. The resulting channel impulse response (CIR) measurements provide both ranging and angular information, which are further processed into 3D point clouds suitable for human pose estimation. The second component is a learning-based pose estimation network that maps radio point clouds to a structured human skeleton. RF-derived point clouds are inherently sparse, noisy, and affected by near-far signal attenuation. To overcome these challenges, RadPose introduces PhysicsAware Sparse Convolution (PASC), Semantic-Guided Spatial Attention (SGSA), and Target-Driven Temporal Attention (TDTA). PASC aggregates information from three dimensions with different physical attributes. SGSA leverages multi-scale semantic features from a feature pyramid network along with learnable queries to suppress background noise and enhance human-centric spatial representations. Further, TDTA adopts an asymmetric frame compression strategy with learnable target-frame queries to aggregate temporal context efficiently while mitigating crossframe feature confusion. The third component is an inverse kinematics (IK) refinement module that enforces kinematic consistency and temporal coherence of the predicted human skeleton. Due to the low spatial resolution and high noise level of RF data, predicted poses may exhibit structural collapse or joint ambiguity, where neighboring joints become indistinguishable. To address this issue, RadPose adopts joint-aware inverse kinematics refinement via a graph convolutional network (GCN) operating on joint-centric embeddings. This module refines the pose estimates by enforcing physical constraints and human body priors, ensuring anatomically plausible and temporally stable pose reconstruction. We built a prototype of RadPose and evaluated it in realistic scenarios. It achieves competitive pose estimation accuracy compared with SOTA radar-based methods, with average estimation errors of 16.2 cm in seen environments and 18.8 cm in unseen environments. Moreover, RadPose maintains stable performance under occlusion, in stark contrast to camera- and LiDAR-based systems. The contributions of this work are as follows: • We present a reliable RF sensing data acquisition method

on individual mmWave communication devices, enabling monostatic sensing without dedicated radar hardware or licensed spectrum. • We address the low spatial resolution of RF sensing with a learning-based pose estimation network that combines semantic-guided spatial attention and target-driven temporal attention, while enforcing kinematic consistency via graph-based refinement. • Extensive experiments demonstrate RadPose’s high pose estimation accuracy in both seen and unseen environments, as well as robust performance under occlusions.

2. Related Work 2.1. Camera- and LiDAR-based Pose Estimation Camera-based 3D human pose estimation has been extensively studied. Existing methods can be classified into two main categories. The first directly predicts 3D joint coordinates from images [11, 26, 40]. The second follows a twostage pipeline, where 2D keypoints are first estimated and then lifted to 3D space [36, 50, 71]. Temporal networks [9, 38] and graph networks [4, 12, 54] have been developed to model time sequences and relationships between joints, so as to improve the estimation accuracy. Recently, transformer-based methods [16, 24, 31, 34, 46, 47, 51, 77, 79] have become the mainstream technique for 3D pose estimation. Although camera-based models have demonstrated remarkable performance, they do not perform well in poor-lighting and occluded scenarios. Compared to camera-based methods, LiDAR-based human pose estimation can accurately localize human joints regardless of illumination. Existing approaches generally fall into two categories: pure LiDAR estimation and multimodal estimation. Pure LiDAR methods employ pointcloud feature extractors such as PointNet and PointNet++ to learn geometric representations [19, 30], with subsequent works enhancing generalization through techniques like knowledge distillation and motion coherence [43, 44], transformer-based global modeling [2, 72], and contextaware modules to better separate human points from noise [43, 75]. Multi-modal methods instead fuse LiDAR with camera data to exploit complementary strengths, as pioneered by HPERL [18], which mitigates RGB depth ambiguity and LiDAR sparsity; later studies extended this strategy to weak supervision and multi-person settings [13, 37, 80]. Nevertheless, LiDAR-based approaches still struggle with occlusion, and the high cost of LiDAR sensors limits their widespread adoption.

2.2. RF-based Pose Estimation MmWave Radar Pose Estimation. MmWave radar has demonstrated its abilities to estimate human 3D pose [20, 32, 61, 68, 76, 81]. Prior work explored diverse radar

data representations including range-Doppler spectra [15], radar point clouds [32, 45, 49, 58, 60, 63], and multi-view heatmap encodings [10, 60, 62]. To address the spatial sparsity and high noise level of radar data, these methods often employ modules such as multi-frame fusion and spatiotemporal attention [1, 23, 60, 62, 64]. Beyond pose skeleton estimation, several works focus on reconstructing full human body meshes [65–67, 70]. Other studies investigate extreme sensing conditions, such as through-wall or aroundcorner scenarios [35, 81]. To enhance the structural correctness, prior work has incorporated various joint refinement strategies like kinematic priors into radar-based pose reconstruction [15, 23, 59, 66, 85]. Despite the promise of radar sensing, its deployment requires dedicated hardware and licensed spectrum. This limitation motivates us to explore fine-grained RF sensing using existing communication infrastructure. WiFi CSI-based Pose Estimation. WiFi CSI-based sensing is a representative ISAC approach that repurposes commodity communication devices for sensing. It has the potential for large-scale deployment because it requires neither new hardware nor licensed spectrum. Early studies demonstrated the feasibility of recovering human poses from WiFi signals using conventional learning-based models [3, 7, 20, 41, 42, 52, 69, 82]. Subsequent studies employed Transformer-based architectures to improve feature representation, expand sensing coverage, and support multiperson pose estimation [55, 68, 73, 83, 84]. However, WiFi CSI-based sensing often generalizes poorly to unseen environments. More recent studies, such as DT-Pose, have therefore focused on cross-environment generalization [5, 6, 21, 22, 39, 56, 57]. Despite these efforts, fundamental generalization challenges remain. This is because CSI measurements obtained in bistatic configurations lack a consistent physical interpretation owing to Tx/Rx hardware impairments and uncertainties in the Tx/Rx geometry [48, 53]. Positioning of RadPose. RadPose investigates monostatic ISAC using a single mmWave communication device for sensing. Compared with radar-based approaches, it requires neither dedicated sensing hardware nor access to licensed spectrum. Compared with WiFi CSI-based methods, its CSI measurements provide physically interpretable features (e.g., ranging, micro-Doppler, and angle of arrival), enabling better generalization to unseen environments. We compare RadPose with other sensing modalities in Table 1.

3. Monostatic ISAC for Data Acquisition MmWave communications have been widely adopted in modern wireless systems, including Wi-Fi (802.11ad) standards and 5G FR2 networks. As illustrated in Figure 1, when a mmWave Wi-Fi or 5G device transmits data packets, its OFDM signals also illuminate the surrounding environment. If the device can simultaneously transmit and

Rx

Tx Comm Device (Transmitter)

Rx

Tx Comm Device (Receiver)

Figure 1. A mmWave communication device operates in fullduplex mode. It is transmitting OFDM signal for data packet transmission; at the same time, its receiver demodulates the OFDM symbols to estimate CIR, which is used for sensing applications.

receive, its receiver can capture the backscattered signals from the surrounding objects and estimate the time-domain channel impulse response (CIR). In such a monostatic setting, these CIR measurements enable precise object ranging, allowing the device to estimate distances to surrounding objects in a manner analogous to monostatic radar. In real systems, mmWave communication devices indeed can operate in full-duplex mode due to their highly directional phased-array antennas and short carrier wavelengths, which together provide sufficient transmitter-receiver isolation.

3.1. CIR Measurement and Interpretation In practice, most mmWave devices use OFDM waveforms for data packet transmission. When a device is transmitting OFDM signals for data transmission, it can receive the backscattered OFDM signals if operating in full-duplex mode (see the transmitter in Figure 1). The received OFDM signals can be used to estimate the channel response, constituting the input data for downstream sensing tasks. Denote K as the number of subcarriers in an OFDM symbol. Denote Yk as the received data symbol and Xk as the transmitted data symbol on OFDM subcarrier k. Then, the channel coefficient of OFDM subcarrier k can be computed by: Hk = Yk /Xk . With the channel coefficient of all subcarriers, the CIR can be calculated
by: h = [h0 , h1 , . . . , hK−1 ] = IFFT [H0 , H1 , . . . , HK−1 ] . When a device operates in full-duplex mode, its transmitter and receiver are co-located on the same device and thus perfectly synchronized in both timing and clock domains. As a result, the CIR measurement can be directly interpreted to object ranging information. Specifically, a large magnitude of CIR coefficient hk indicates a high probabilkc ity of an object present at a distance of 2W , where c is light speed and W is OFDM signal bandwidth. This ranging capability fundamentally differs RadPose from WiFi CSIbased sensing methods.

3.2. Radio Point Cloud Generation Commercial mmWave communication devices are widely equipped with phased-array antennas for signal beamforming. The phased-array antennas can be leveraged for angular estimation based on the CIR measurement, generating the point cloud for downstream tasks. Consider a

z

voxel

Rx

Tx

x

ϕ θ

y

Figure 2. Angular estimation on a mmWave device.

voxel in the 3D space as shown in Figure 2. Denote θ as its azimuth angle and ϕ as its elevation angle. The direction of this voxel can then be written as: ⃗u(θ, ϕ) = [cos(θ) cos(ϕ), cos(θ) sin(ϕ), sin(ϕ)]. Denote Atx as the set of Tx antennas and Arx as the set of Rx antennas. Denote p̃⃗i = (x̃i , 0, z̃i ) as the coordinate of Tx antenna i ∈ Atx and p̄⃗j = (x̄j , 0, z̄j ) as the coordinate of Rx antenna j ∈ Arx . When antenna i transmits and antenna j receives signal, the additional signal travel distance in the direction of (θ, ϕ) can be computed by: ⟨⃗u(θ, ϕ), p̃⃗i + p̄⃗j ⟩. Therefore, the complex weight associated with the antenna pair (i, j) can be written as: αi,j (θ, ϕ) = exp − j

2π ⟨⃗u(θ, ϕ), p̃⃗i + p̄⃗j ⟩ , λ

(1)

where λ is wavelength of carrier signal and the first j is the imaginary unit. Denote hi,j = [hi,j,0 , hi,j,1 , . . . , hi,j,K−1 ] as the CIR measurement when RadPose uses antenna i for signal transmission and antenna j for signal reception. Then, the reflected signal from voxel (k, θ, ϕ) can be estimated by: X X s(k, θ, ϕ) = αi,j (θ, ϕ) · hi,j,k , (2) i∈Atx j∈Arx kc where k represents the voxel distance (i.e., r = 2W ). The magnitude of s(k, θ, ϕ) represents the likelihood of an object at this voxel position, while the temporal phase variation of s(k, θ, ϕ) reveals the object’s velocity, if present. For radio point cloud generation, let K be the set of range indices, Θ the set of azimuth angles, and Φ the set of elevation angles. A radio point cloud frame can be obtained by evaluating Eq. (2) over these domains, which we denote as  S = s(k, θ, ϕ) | k ∈ K, θ ∈ Θ, ϕ ∈ Φ . (3)

Appendix A provides more details on the radio point cloud generation and illustration.

4. Method RadPose takes consecutive radio point cloud frames as input to estimate the 3D human pose. Figure 3 illustrates the network pipeline. It consists of two main components: (i) a pose estimator and (ii) an inverse kinematics module. The pose estimator first voxelizes the noisy radio point clouds, followed by a physics-aware encoder and spatio-temporal

attention modules. A heatmap prediction head then produces joint heatmaps by learning discriminative spatial and temporal representations from the radio point cloud sequences. The inverse kinematics module employs a GCN, where the predicted heatmaps serve as node features. By explicitly modeling human kinematic constraints, this module refines the pose skeleton generated by the pose estimator to ensure anatomically consistent predictions.

4.1. Pose Estimator Voxelization. The radio point clouds in Eq. (3) are represented in spherical coordinates. Within the spherical region of interest, we partition the space into voxel grids by uniformly discretizing the radial distance, azimuth angle, and elevation angle. For each radio frame, we first filter out points whose amplitude values are below a threshold, which is empirically set. Points with low intensity are likely caused by circuit noise or multipath reflections, and thus do not carry object information. The remaining radio points are then discretized into a 3D spherical voxel grid. The resulting set of voxels at frame t is defined as: Vt = {(pi , log(ci ), ai , ϕi ) | ci > 0, i ∈ I}, where t denotes the frame index, I is the set of valid voxel indices, pi is the center coordinate of voxel i, ci is the number of radio points within voxel i, ai is the average amplitude of the points within voxel i, and ϕi is the average phase of the points within voxel i. In practice, the majority of voxels are empty. We thus remove empty voxels to reduce the data dimensionality for subsequent processing. Physics-Aware Encoder. In the spherical coordinate system, the spatial extent covered by a voxel increases with radial distance (i.e., the near-far effect). Conventional sparse convolution treats each dimension of voxels as isotropic, which is appropriate in Cartesian space where voxel extents are uniform. However, in spherical space, dimensions split into radial and angular ones, each exhibiting distinct physical characteristics. Using conventional sparse convolution tends to cause feature entanglement and suboptimal spatial modeling. To address this, RadPose performs anisotropic convolution on each individual radio frame over the three branches for dimension-specific voxel feature extraction. Then, it performs physics-aware aggregation through a linear combination of the resulting embeddings, followed by positional encoding along the spatial and temporal dimensions. We denote Xt as the output encoded token from this module, corresponding to the voxelized radio frame Vt . Details of the design can be found in Appendix B. Semantic-Guided Spatial Attention. Due to the sparsity and high noise level of RF-derived data, a large number of background voxels overwhelm the informative human voxels. In addition, the scale diversity of human movements is challenging for Transformers to model. As a result, spatial

Radio Point Cloud Frames ……

……

……

……

……

Physics-Aware Sparse Convolution Positional Encoding Token Embedding

Temporal Encoding …… …… …… ……

Physics-Aware Encoder

Refined Joint Coordinates

Coarse Joint Coordinates Root Joint

…… …… …… ……

…… …… …… ……

GCN

FK

Target Frame Non-Target Frame

Attention-based Down-sampling ……

Voxels

…… …… …… ……

…… …… …… ……

Feature Pyramid Network

Voxelization

Spatial Tokens

Token Embedding

Attention-based Down-sampling

…… …… …… ……

Temporal Transformer Cross Attention

…… …… …… ……

Spatial Attention Query …… …… …… ……

…… …… …… ……

Cross Attention

Spatial Tokens

Target-Driven Temporal Attention

Semantic-Guided Spatial Attention

Pose Estimator

CNN Heatmap Head

……

Joint Query

Heatmap Joint Coordinates

Joint-Aware Inverse Kinematic Refinement

Figure 3. Overview of RadPose. Physics-aware encoder and spatio-temporal attention aggregate spatio-temporal features from radio point clouds, and extract target frame’s feature for joint estimation. Joint-aware inverse kinematic refinement then estimates initial heatmaps and applies structural constraints to generate the refined 3D coordinates.

attention alone is insufficient. To address this, RadPose first employs a Feature Pyramid Network (FPN) to extract multiscale features, which provide more detailed semantics indicating region of interest for transformer and remain robust for human detection during movements. Then, it adopts learnable queries to provide additional semantics for token downsampling. This operation enhances useful tokens for pose estimation and reduces transformer memory consumption. The spatial features of downsampled tokens, which we denote as Xst , are extracted by spatial self-attention. Target-Driven Temporal Attention. Temporal attention across multiple consecutive frames is essential for improving estimation accuracy. Accordingly, RadPose leverages T consecutive radio frames to estimate one pose. Conventional cross-frame attention treats all T frames equally, resulting in a high-dimensional token space that increases computational complexity for the network. To address this issue, RadPose designates one of the T frames as the target frame and the others as non-target frames. Full tokens are retained only for the target frame, while tokens from non-target frames are compressed into high-semantic tokens. This design reduces background noise from nontarget frames while still providing complementary semantic cues from neighboring frames, enabling more focused temporal modeling. Additionally, directly decoding pose from the aggregated feature tends to cause information confusion. To avoid this, RadPose employs a target frame decoding strategy by introducing learnable target-frame queries. The learnable queries explicitly retrieve pose-relevant information corresponding to the target frame, ensuring the final representation preserves the frame-specific structure while benefiting from temporal aggregation. Specifically, the operation



h

i h

i

is Xout = MHA Qtar , Xstar , Xsntar↓ , Xstar , Xsntar↓ , where Qtar is the target frame query; Xsntar↓ is the downsampled feature of Xsntar , Xsntar = {Xst | 1 ≤ t ≤ T, t ̸= tar}. Heatmap Head. Given the large number of voxels, the pose estimation network tends to have sparse gradients as the loss function may not capture every dimension of the input data. To address this issue, RadPose adopts a 2.5D pose heatmap in the (r, θ) and (r, ϕ) planes as the output of the model, following the principle presented in [25]. It provides dense information that can significantly reduce the gradient sparsity. To generate the 2.5D pose heatmaps, RadPose employs a conv/deconv head to reshape the target-frame tokens Xout into a spatial feature map. The feature map is then progressively upsampled using multiple transposed convolutions to produce the range-azimuth heatmaps RA and elevation maps E. For each joint, the coordinate is obtained via soft-argmax on the heatmaps, while the confidence is estimated from the entropy of the heatmap distribution. Details of heatmap head are given in Appendix C.

4.2. Joint-Aware Inverse Kinematic Refinement Poses generated by the above model frequently suffer from inconsistent limb lengths, implausible joint angles, and locally collapsed structures. This occurs because, although the transformer backbone is effective at capturing contextual relationships among spatial and temporal tokens, it does not explicitly encode or enforce human body kinematics. As a result, the predicted joint locations may satisfy feature-level correlations but violate fundamental anatomical constraints. To address this issue, RadPose employs a GCN for inverse kinematic refinement. By treating joints as nodes and bones as edges, the GCN explicitly enforces relational con-

sistency among adjacent joints and propagates structural information across the kinematic tree. Joint Embedding. Let P̂ = [p̂1 , . . . , p̂J ] be the joint coordinates decoded from the heatmap head. For each joint j, we construct a joint query by combining a learnable embedding with the decoded joint coordinate. Each joint query attends to the transformer tokens to produce a joint-centric feature: hj = MHA(qj , Xout , Xout ), qj = ej + W̃pos p̂j , (4) where ej is the learnable embedding of joint j, and W̃pos projects the coordinate into the feature space. hj encodes pose-relevant information for joint j. GCN for Inverse Kinematics. RadPose treats the human skeleton as a kinematic graph. The initial node feature of joint j is formed by fusing the joint-centric fea(0) ture hj , decoded coordinate p̂j , and confidence cj : gj = [hj , p̂j , cj ]. The node features are then propagated on the kinematic graph:
G(ℓ+1) = σ ÃG(ℓ) W(ℓ) , ℓ = 0, . . . , Lg −1, (5) where G(ℓ) stacks all node features at layer ℓ, Ã ∈ RJ×J is the normalized confidence-aware adjacency matrix. W(ℓ) is the learnable weight matrix, σ(·) is the activation function. From the final graph features, RadPose predicts a direction vector dj , a bone-length scale sj , and a residual root offset ∆pr using a model parameterized by ψ: (L )
[dj , sj , ∆pr ] = ψ gj g . (6) Forward Kinematics. In the final step, RadPose selects one joint as the root joint for pose construction, Let r be its index. The refined root joint coordinate is computed as pr = p̂r + ∆pr . For each non-root joint j, its position is reconstructed recursively by forward kinematics: pj = pπ(j) + (lj sj )dj , where j ̸= r; π(j) denotes the parent joint of j; lj is the predefined canonical bone length; and sj is the predicted scale factor. Finally, a small residual correction is added to improve flexibility: pik j = pj + ∆pj , which constitutes the final coordinate of joint j.

4.3. Supervised Training RadPose employs a two-stage supervised training strategy. It first trains the pose estimator model and then trains GCN for the inverse kinematic module. Pose Estimator Training. The loss function of training the pose estimator is a weighted summation of heatmap loss Lmap , coordinate loss Lxyz , bone length loss Llen and joint angle loss Lang , where Lxyz , Llen and Lang are the smoothL1 losses between predicted value and ground-truth labels. In particular, Lmap is the most important loss for this training. It alleviates the sparse gradient issue inherent in direct coordinate regression, as it supervises a spatial distribution rather than a single point estimate. It is

Rx Phased-array Antenna Tx Phased-array Antenna Commercial mmWave Transceiver (Sivers EVK06002) Depth Camera (Intel RealSense D455)

12V 9AH SLA Battery

AMD/Xilinx RFSoC 4x2 FPGA

RGB Camera (Unused)

Figure 4. Hardware components of RadPose.

 P 1 β defined as: Lmap = − |Ω| u∈Ω αyu (1 − ŷu ) log ŷu +  (1 − yu )ŷuβ log(1 − ŷu ) , where u indexes all spatial locations on the predicted heatmap and Ω is its complete set, yu ∈ [0, 1] and ŷu ∈ [0, 1] denote the ground-truth and predicted heatmap values at location u, respectively, and α and β are the focusing parameters that balance positive and negative samples and emphasize hard examples. GCN Training. After training pose estimator, we then train the GCN for the inverse kinematic module. During this stage, the GCN is trained with full scale learning rate, while the parameters of pose estimator are trained with a small learning rate to preserve the representations learned by the pose estimator while allowing it to be slightly refined to better cooperate with the inverse kinematics module. In this stage, the loss of root joint Lroot is added to penalize the undesired topology of human pose.

5. Implementation Hardware Setup. We built a prototype of RadPose, as shown in Figure 4, to evaluate its performance in realistic environments. The hardware integrates two sensing components: (i) a custom-designed mmWave 5G/Wi-Fi OFDM communication device that supports full-duplex operation for monostatic sensing, and (ii) a depth camera (Intel RealSense D455) that provides ground-truth pose annotations. The mmWave device operates at 60 GHz with a bandwidth of 1.2288 GHz and 16 Tx-Rx antenna elements. It offers an sensing range of 6 m, covering a field of view of 120◦ horizontally and 60◦ vertically. Importantly, the device supports simultaneous data transmission and sensing (see Appendix E). The system was built using commercial off-the-shelf RF components. Network Parameters. RadPose voxelizes the input into 20,480 voxels, and five consecutive radio frames are used for pose estimation. Three physics-aware sparse convolutions encode each voxel into a 512-dimensional token, followed by an attention-based downsampling that reduces the tokens to 2,048. For inverse kinematics, the 6th joint is selected as the root, and a three-layer GCN with 512 hidden dimensions is used as the IK head. More details are provided in Appendix D.

Table 2. Quantitative comparison of RadPose against SOTA RF sensing methods. All MPJPE values are absolute errors without root alignment. [S./U. denotes seen/unseen scenes, P.C. denotes radio point clouds, B.W. denotes Bandwidth, ‘–’ denotes ‘Not Available’.] Methods

Input

# of scenes (S./U.)

# of activities

Waveform

B.W. (GHz)

∆R (cm)

MPJPE (S./U.)

PCK10 (S./U.)

MPJPE/∆R (S./U.)

Person-in-WiFi 3D [68] GraphPose-Fi [5]

CSI CSI

3/1 3/1

8 14

OFDM OFDM

– –

– –

9.2/62.6 16.1/30.2

–/– 33.3/2.7

–/– –/–

mmMesh [65] MvDoppler [10]

P.C. Range/Doppler

1/1 1/0

8 1

FMCW FMCW

3.9 1.0

3.8 15.0

2.2/5.1 9.7/–

–/– –/–

0.58/1.34 0.64/–

mmDiff [17] Wave2Body [33] RadPose

P.C. (Our dataset) P.C. (Our dataset) P.C. (Our dataset)

17/4 17/4 17/4

Unconstrained Unconstrained Unconstrained

OFDM OFDM OFDM

1.2 1.2 1.2

12.2 12.2 12.2

25.1/61.4 35.5/38.9 16.2/18.8

16.5/0.5 10.9/10.1 36.3/29.8

2.06/5.04 2.91/3.18 1.33/1.54

Dataset Details. We collected 183,497 radio frames with the ground-truth pose annotations of 11 subjects in 21 scenes. The ground-truth annotations are derived from corresponding depth camera frames with pretrained RTMPose model [27]. Unlike [5, 10, 65], where subjects are restricted to predefined movements, our dataset covers the subjects without motion sequences, fixed motion amplitudes, or designated activity locations. Of the 21 scenes, 75% data of 17 scenes (120,174 frames) are used for model training. The remaining 25% data (39,400 frames) and holdout 4 scenes (23,923 frames) are reserved for evaluation. We additionally collected 14,209 frames with constrained movements to better evaluate influence of activity types.

6. Experiments 6.1. Main Results Two metrics are used for evaluation: Mean Per Joint Position Error (MPJPE/cm) andPPercentage of Keypoints J (PCKδ /%). MPJPE = J1 j=1 ∥p̂j − pj ∥2 , where J is the total number of joints, p̂j and pj are estimated of joint j. PCKδ =
PJ and ground-truth positions 1 I ∥p̂ − p ∥ < δ , where I(·) is the indicator j j j=1 2 J function, δ is a joint error distance threshold. Both metrics are based on absolute 3D joint positions without alignment. Quantitative Results. We first evaluate RadPose’s quantitative performance and its ability to generalize to unseen environments. Table 2 presents our experimental results. The results show that RadPose maintains robust performance on held-out test scenes. Notably, the error of root joint (i.e., the 6th joint) is 12.1 cm, which indicates the precise localization ability of our model. Comparison with SOTA. We compare RadPose with SOTA radar-based and WiFi CSI-based pose estimation methods. MmMesh [65], Mv-Doppler [10], mmDiff [17], and Wave2Body [33] are representative FMCW radar-based methods, while Person-in-WiFi 3D [68] and GraphPoseFi [5] are representative WiFi CSI-based methods. In RF sensing, bandwidth W determines the range resolution, c , and fundamentally affects pose estimation ac∆R = 2W curacy. To account for differences in bandwidth, we use

the normalized MPJPE, MPJPE ∆R , as the comparison metric. Table 2 reports the comparison results. Compared with the WiFi CSI-based methods (Person-inWiFi and GraphPose-Fi), RadPose performs slightly worse in seen environments but substantially better in unseen environments. This improved generalization likely stems from its monostatic sensing geometry and physically interpretable signal features, whereas WiFi CSI methods rely more heavily on data-driven feature learning from bistatic measurements. Compared with the radar-based methods (MmMesh and Mv-Doppler), RadPose has slightly higher normalized errors: 1.33 versus 0.58 in seen scenes and 1.54 versus 1.34 in unseen scenes. This gap may arise for three reasons. First, MmMesh and Mv-Doppler use dedicated FMCW radars, which are generally more robust to noise than OFDM-based sensing. Second, their evaluation settings are more limited: MmMesh considers two scenes and Mv-Doppler only one, whereas RadPose is evaluated in 17 seen and four unseen scenes. Third, they primarily evaluate walking activities, while RadPose includes more unconstrained activities. Despite this performance gap, RadPose requires neither dedicated sensing hardware nor licensed spectrum and therefore targets different application domains. To evaluate our network pipeline, we also test mmDiff and Wave2Body on our dataset. As shown in Table 2, RadPose substantially outperforms both methods, demonstrating the benefits of PASC, SGSA, TDTA, and jointaware IK refinement. Qualitative Results. Figure 5 presents a visual comparison of our pose estimation. The first row are scenes used in training, and the second row are unseen scenes. From visualization, most errors come from wrists and ankles, which are difficult for RadPose to detect due to their small physical size. More pose visualization is provided in Appendix F-G, and a demo video is available on the project website. While most experiments demonstrate accurate pose estimation, we also observe some failure cases in the results. Some representative examples are provided in Appendix H. Impact of Different Activities. Table 3 reports the performance of RadPose across different activities. RadPose

Table 4. Occlusion robustness.

Table 5. Effectiveness of IK refinement.

Occlusion MPJPE PCK10 None Cloth Styrofoam

12.5 15.0 13.4

42.7 29.2 38.2

IK

MPJPE PCK10

w/o w

17.0 16.2

14.0 36.3

Table 6. Subject generalization with and without IK refinement. NEA/L/B denotes normalized error for arms/legs/body (%) w/o IK Subjects MPJPE↓ Seen Unseen

Figure 5. Qualitative performance of RadPose in different scenes. The green skeleton represents the ground-truth labels derived from depth camera while the blue skeleton is the results of RadPose.

achieves the highest accuracy for walking, likely because its regular and periodic motion patterns facilitate temporal modeling. Squatting and standing are the most challenging, as compact body configurations bring multiple joints close together, increasing self-occlusion and ambiguity in the reflected signals. Nevertheless, RadPose maintains consistent performance on mixed activities, demonstrating its robustness across diverse motions. Table 3. RadPose’s performance across different activities. Activities

Walking

Waving

Sit/Stand

Squat/Stand

Mixed

MPJPE PCK10

13.17 46.9

18.72 29.7

16.07 39.1

21.56 18.4

16.21 36.3

Robustness to Occlusion. Compared to the camera- and LiDAR-based pose estimation, one salient advantage of RadPose is its robustness to occlusions. To demonstrate this advantage, we conducted experiments using cloth and foam to completely cover the mmWave device’s antennas but leave the depth camera uncovered. Table 4 reports the experimental results. RadPose’s performance degrades slightly under these occlusions. In particular, MPJPE remains within 15 cm and the PCK10 remains above 29%, indicating RadPose maintains a robust performance when the antennas are completely covered. This reveals that RadPose can be used to complement camera and LiDAR in challenging scenarios and its potential for covert surveillance. More occlusion results are given in Appendix H.

6.2. Ablation Study Joint-Aware IK Refinement. The IK refinement is the key component of RadPose. To evaluate its effectiveness, we measure the pose estimation errors of RadPose with and without the IK module. Table 5 presents the results. The IK module reduces the estimation error from 17.0 cm to 16.2 cm, with PCK10 improved 22.3%. This significant improvement can be attributed to the GCN within the IK mod-

14.66 17.18

w/ IK

NEA/L/B ↓

MPJPE↓

NEA/L/B ↓

33.91/18.48/13.21 38.66/21.82/16.19

13.91 16.30

30.80/18.16/12.88 35.41/21.35/15.59

ule, which effectively encodes skeletal nodes and edges to capture human pose geometry. To evaluate the generalization of our IK refinement module, we measure RadPose’s performance across subjects. The model was trained on seven-subject dataset with four held out. Table 6 presents the results. The IK refinement leads to pronounced improvements in estimation accuracy across all subjects. Other Components. More ablation studies of RadPose’s key components are given in Appendix I.

7. Conclusion and Discussion In this paper, we presented RadPose, a monostatic ISAC framework that enables individual mmWave communication devices to perform 3D human pose estimation. The key innovations of RadPose include (i) a cross-frame transformer for representing mmWave point clouds and (ii) a novel inverse kinematics method for mmWave sensing that leverages body-topology constraints to improve pose estimation accuracy. Experimental results validate the effectiveness of RadPose in realistic environments. RadPose demonstrates the feasibility of using OFDMbased communication signals for fine-grained perception with a single mmWave device. However, the current system has two limitations. First, RadPose supports only singleperson pose estimation. Future work is needed to extend the learning framework to multi-person scenarios. Appendix K presents a plausible extension of RadPose for such scenarios. Second, RadPose is not currently compatible with commercial Wi-Fi or 5G communication devices because they do not support full-duplex operation. Integrating RadPose into future communication infrastructure will require changes to communication standards, as well as software and firmware upgrades to communication devices.

Acknowledgments We sincerely thank the reviewers for their insightful comments. This work is supported in part by the NSF under Award No. ECCS-2225337 and ECCS-2434001.

References [1] Aakriti Adhikari and Sanjib Sur. Argosleep: Monitoring sleep posture from commodity millimeter-wave devices. In IEEE INFOCOM 2023-IEEE Conference on Computer Communications, pages 1–10. IEEE, 2023. 3 [2] Xiaoqi An, Lin Zhao, Chen Gong, Jun Li, and Jian Yang. Pre-training a Density-Aware Pose Transformer for Robust LiDAR-based 3D Human Pose Estimation, 2024. 2 [3] Xiaojun Bi, Shuo Li, Changdong Yu, and Yu Zhang. A novel approach using convolutional transformer for massive mimo csi feedback. IEEE Wireless Communications Letters, 11(5): 1017–1021, 2022. 3 [4] Yujun Cai, Liuhao Ge, Jun Liu, Jianfei Cai, Tat-Jen Cham, Junsong Yuan, and Nadia Magnenat Thalmann. Exploiting spatial-temporal relationships for 3d pose estimation via graph convolutional networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2272–2281, 2019. 2 [5] Jichao Chen, YangYang Qu, Ruibo Tang, and Dirk Slock. Graph-based 3d human pose estimation using wifi signals. In ICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 19992–19996. IEEE, 2026. 3, 7 [6] Yang Chen and Jingcai Guo. Dt-pose: Towards robust and realistic human pose estimation using wifi signals. Edge Intelligence and Systems, 1(1):2, 2026. 3 [7] Yi-Chung Chen, Zhi-Kai Huang, Lu Pang, Jian-Yu JiangLin, Chia-Han Kuo, Hong-Han Shuai, and Wen-Huang Cheng. Seeing the unseen: Wifi-based 2d human pose estimation via an evolving attentive spatial-frequency network. Pattern Recognition Letters, 171:21–27, 2023. 3 [8] Yu Cheng, Bo Yang, Bo Wang, Wending Yan, and Robby T Tan. Occlusion-aware networks for 3d human pose estimation in video. In Proceedings of the IEEE/CVF international conference on computer vision, pages 723–732, 2019. 1 [9] Yu Cheng, Bo Yang, Bo Wang, and Robby T Tan. 3d human pose estimation using spatio-temporal networks with explicit occlusion training. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10631–10638, 2020. 2 [10] Jaeho Choi, Soheil Hor, Shubo Yang, and Amin Arbabian. Mvdoppler-pose: Multi-modal multi-view mmwave sensing for long-distance self-occluded human walking pose estimation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 27750–27759, 2025. 1, 3, 7, 15 [11] Sungho Chun, Sungbum Park, and Ju Yong Chang. Learnable human mesh triangulation for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2850–2859, 2023. 2 [12] Hai Ci, Chunyu Wang, Xiaoxuan Ma, and Yizhou Wang. Optimizing network structure for 3d human pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2262–2271, 2019. 2 [13] Peishan Cong, Yiteng Xu, Yiming Ren, Juze Zhang, Lan Xu, Jingya Wang, Jingyi Yu, and Yuexin Ma. Weakly supervised 3d multi-person pose estimation for large-scale scenes based

on monocular camera and single lidar. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 461–469, 2023. 2 [14] Viviana Crescitelli, Atsutake Kosuge, and Takashi Oshima. Poison: Human pose estimation in insufficient lighting conditions using sensor fusion. IEEE Transactions on Instrumentation and Measurement, 70:1–8, 2020. 1 [15] Wen Ding, Zhongping Cao, Jianxiong Zhang, Rihui Chen, Xuemei Guo, and Guoli Wang. Radar-based 3d human skeleton estimation by kinematic constrained learning. IEEE Sensors Journal, 21(20):23174–23184, 2021. 3 [16] Moritz Einfalt, Katja Ludwig, and Rainer Lienhart. Uplift and upsample: Efficient 3d human pose estimation with uplifting transformers. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2903–2913, 2023. 2 [17] Junqiao Fan, Jianfei Yang, Yuecong Xu, and Lihua Xie. Diffusion model is a good pose estimator from 3d rf-vision. In European Conference on Computer Vision, pages 1–18. Springer, 2024. 7 [18] Michael Fürst, Shriya TP Gupta, René Schuster, Oliver Wasenmüller, and Didier Stricker. Hperl: 3d human pose estimation from rgb and lidar. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 7321–7327. IEEE, 2021. 2 [19] Liuhao Ge, Zhou Ren, and Junsong Yuan. Point-to-point regression pointnet for 3d hand pose estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 2 [20] Jiaqi Geng, Dong Huang, and Fernando De la Torre. Densepose from wifi. arXiv preprint arXiv:2301.00250, 2022. 2, 3 [21] Toan D Gian, Dung T Tran, Quoc-Viet Pham, Le-Nam Tran, and Van-Dinh Nguyen. Multi-modal human pose estimation: A wi-fi-driven approach with adaptive kernel selection. IEEE Transactions on Artificial Intelligence, 2025. 3 [22] Yanling Hao, Arumugam Nallanathan, and Yuanwei Liu. Latent diffusion-inspired domain-invariant human pose estimation from wifi signals. In 2025 IEEE/CIC International Conference on Communications in China (ICCC), pages 1–6. IEEE, 2025. 3 [23] Shuting Hu, Siyang Cao, Nima Toosizadeh, Jennifer Barton, Melvin G Hector, and Mindy J Fain. mmpose-fk: A forward kinematics approach to dynamic skeletal pose estimation using mmwave radars. IEEE sensors journal, 24(5): 6469–6481, 2024. 1, 3 [24] Yunlong Huang, Junshuo Liu, Ke Xian, and Robert Caiming Qiu. Posemamba: Monocular 3d human pose estimation with bidirectional global-local spatio-temporal state space model. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3842–3850, 2025. 1, 2 [25] Umar Iqbal, Pavlo Molchanov, Thomas Breuel Juergen Gall, and Jan Kautz. Hand pose estimation via latent 2.5 d heatmap regression. In Proceedings of the European conference on computer vision (ECCV), pages 118–134, 2018. 5, 14 [26] Karim Iskakov, Egor Burkov, Victor Lempitsky, and Yury Malkov. Learnable triangulation of human pose. In Proceed-

ings of the IEEE/CVF international conference on computer vision, pages 7718–7727, 2019. 2 [27] Tao Jiang, Peng Lu, Li Zhang, Ningsheng Ma, Rui Han, Chengqi Lyu, Yining Li, and Kai Chen. Rtmpose: Realtime multi-person pose estimation based on mmpose. arXiv preprint arXiv:2303.07399, 2023. 7 [28] Hakpyeong Kim, Juwon Hong, Eunseong Song, Taehoon Hong, and Jun-Ki Choi. A multi-dimensional failure factor affecting the on-site adoption of construction robots. Developments in the Built Environment, page 100841, 2025. 1 [29] Sohyun Lee, Jaesung Rim, Boseung Jeong, Geonu Kim, Byungju Woo, Haechan Lee, Sunghyun Cho, and Suha Kwak. Human pose estimation in extremely low-light conditions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 704–714, 2023. 1 [30] Jialian Li, Jingyi Zhang, Zhiyong Wang, Siqi Shen, Chenglu Wen, Yuexin Ma, Lan Xu, Jingyi Yu, and Cheng Wang. Lidarcap: Long-range marker-less 3d human motion capture with lidar point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20502–20512, 2022. 2 [31] Wenhao Li, Hong Liu, Runwei Ding, Mengyuan Liu, Pichao Wang, and Wenming Yang. Exploiting temporal contexts with strided transformer for 3d human pose estimation. IEEE Transactions on Multimedia, 25:1282–1293, 2022. 2 [32] Wei Li, Wen Lei, Kun Shi, Zhiguo Shi, Yong Wang, and Jinhai Zhou. mmskeleton: 3d human skeleton estimation using millimeter wave radar sparse point clouds. In 2024 IEEE/CIC International Conference on Communications in China (ICCC), pages 307–312. IEEE, 2024. 2, 3 [33] Bo Liang, Chen Gong, Wei Gao, and Chenren Xu. Wave2body: Rethinking mmwave human pose estimation as radar-to-body token translation. arXiv preprint arXiv:2607.18875, 2026. 7 [34] Jiajie Liu, Mengyuan Liu, Hong Liu, and Wenhao Li. Tcpformer: Learning temporal correlation with implicit pose proxy for 3d human pose estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5478– 5486, 2025. 2 [35] Jiaxi Lv, Guiyun Fan, Xinyue Fu, Jiahui Sun, Rong Ding, and Haiming Jin. mmwave-based relay reflector reconstruction for lidar-free around-corner human sensing. In IEEE INFOCOM 2025-IEEE Conference on Computer Communications, pages 1–10. IEEE, 2025. 3 [36] Soroush Mehraban, Vida Adeli, and Babak Taati. Motionagformer: Enhancing 3d human pose estimation with a transformer-gcnformer network. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 6920–6930, 2024. 2 [37] Zhiyu Pan, Zhicheng Zhong, Wenxuan Guo, Yifan Chen, Jianjiang Feng, and Jie Zhou. LiCamPose: Combining Multi-View LiDAR and RGB Cameras for Robust Singletimestamp 3D Human Pose Estimation. 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2484–2494, 2025. 2 [38] Dario Pavllo, Christoph Feichtenhofer, David Grangier, and Michael Auli. 3d human pose estimation in video with tem-

poral convolutions and semi-supervised training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7753–7762, 2019. 2 [39] Yanyi Qu, Haoyang Ma, and Wenhui Xiong. Multiformer: A multi-person pose estimation system based on csi and attention mechanism. IEEE Internet of Things Journal, 2026. 3 [40] N Dinesh Reddy, Laurent Guigues, Leonid Pishchulin, Jayan Eledath, and Srinivasa G Narasimhan. Tessetrack: End-toend learnable multi-person articulated 3d pose tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15190–15200, 2021. 2 [41] Yili Ren, Zi Wang, Sheng Tan, Yingying Chen, and Jie Yang. Winect: 3d human pose tracking for free-form activity using commodity wifi. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 5(4):1–29, 2021. 3 [42] Yili Ren, Zi Wang, Yichao Wang, Sheng Tan, Yingying Chen, and Jie Yang. Gopose: 3d human pose estimation using wifi. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 6(2):1–25, 2022. 3 [43] Yiming Ren, Xiao Han, Yichen Yao, Xiaoxiao Long, Yujing Sun, and Yuexin Ma. Livehps++: Robust and coherent motion capture in dynamic free environment. In European Conference on Computer Vision, pages 127–144. Springer, 2024. 2 [44] Yiming Ren, Xiao Han, Chengfeng Zhao, Jingya Wang, Lan Xu, Jingyi Yu, and Yuexin Ma. Livehps: Lidar-based scenelevel human pose and shape estimation in free environment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1281–1291, 2024. 1, 2 [45] Arindam Sengupta, Feng Jin, Renyuan Zhang, and Siyang Cao. mm-pose: Real-time human skeletal posture estimation using mmwave radars and cnns. IEEE sensors journal, 20 (17):10032–10044, 2020. 3 [46] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Shanshe Wang, Siwei Ma, and Wen Gao. P-stmo: Pre-trained spatial temporal many-to-one model for 3d human pose estimation. In European Conference on Computer Vision, pages 461–478. Springer, 2022. 2 [47] Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Zhao Wang, Kai Han, Shanshe Wang, Siwei Ma, and Wen Gao. Diffusion-based 3d human pose estimation with multihypothesis aggregation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14761– 14771, 2023. 2 [48] Kunzhe Song, Qijun Wang, Shichen Zhang, and Huacheng Zeng. Siwis: Fine-grained human detection using single wifi device. In Proceedings of the 30th Annual International Conference on Mobile Computing and Networking, pages 1439– 1454, 2024. 3 [49] Kunzhe Song, Geo Jie Zhou, Xiaoming Liu, and Huacheng Zeng. Rascene: High-fidelity 3d scene imaging with mmwave communication signals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 36454–36463, 2026. 3

[50] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5693–5703, 2019. 1, 2 [51] Zhenhua Tang, Zhaofan Qiu, Yanbin Hao, Richang Hong, and Ting Yao. 3d human pose estimation with spatiotemporal criss-cross attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4790–4799, 2023. 2 [52] Fei Wang, Sanping Zhou, Stanislav Panev, Jinsong Han, and Dong Huang. Person-in-wifi: Fine-grained person perception using wifi. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 5451–5460. IEEE, 2019. 3 [53] Fei Wang, Tingting Zhang, Wei Xi, Han Ding, Ge Wang, Di Zhang, Yuanhao Cui, Fan Liu, Jinsong Han, Jie Xu, and Tony Xiao Han. A survey on wi-fi sensing generalizability: Taxonomy, techniques, datasets, and future research prospects. IEEE Communications Surveys & Tutorials, 28: 5227–5266, 2026. 3 [54] Jingbo Wang, Sijie Yan, Yuanjun Xiong, and Dahua Lin. Motion guided 3d pose estimation from videos. In European conference on computer vision, pages 764–780. Springer, 2020. 2 [55] Liming Wang, Lingchao Guo, Zhaoming Lu, Xiangming Wen, and Shuang Zhou. Wilink: Link selection-based 3d human pose estimation using commodity wi-fi. In 2023 IEEE Wireless Communications and Networking Conference (WCNC), pages 1–7. IEEE, 2023. 3 [56] Weili Wang, Weixiong Zhang, Linjun Zhao, Benying Tan, Huijun Wu, and Huakun Huang. Efficientwipose: A lightweight and high-accuracy wifi-based human pose estimation system for energy-efficient smart buildings. In Cold Climate HVAC Conference, pages 326–333. Springer, 2025. 3 [57] Weili Wang, Weixiong Zhang, Lingjun Zhao, Jiamin Yao, Qinglin Yang, and Huakun Huang. Efficient wifi-based human pose estimation via lightweight model design with knowledge distillation. In 2025 IEEE 18th International Symposium on Embedded Multicore/Many-core Systems-onChip (MCSoC), pages 700–705. IEEE, 2025. 3 [58] Yujun Wang, Fabian Seguel, Stefan Hägele, Driton Salihu, and Eckehard Steinbach. Points: mmwave radar-based pointnet and transformer network for human skeleton estimation in privacy-preserving rehabilitation. In IEEE EUROCON 2025-21st International Conference on Smart Technologies, pages 1–6. IEEE, 2025. 3 [59] Ta-Wei Wu, Shih-Hau Fang, Hsiao-Chun Wu, Guannan Liu, and Kun Yan. Robust skeletal-graph reconstruction using mmwave radar and its application for human-activity recognition. IEEE Journal of Selected Areas in Sensors, 2025. 3 [60] Yingxiao Wu, Zhongmin Jiang, Haocheng Ni, Changlin Mao, Zhiyuan Zhou, Wenxiang Wang, and Jianping Han. mmhpe: Robust multi-scale 3d human pose estimation using a single mmwave radar. IEEE Internet of Things Journal, 2024. 3

[61] Jiaren Xiao, Zhiguo Chen, and Bing Luo. A novel two-stage net for human body region segmentation via centimeterwave radar. In 2023 3rd International Conference on Neural Networks, Information and Communication Engineering (NNICE), pages 104–107. IEEE, 2023. 2 [62] Chunyang Xie, Dongheng Zhang, Zhi Wu, Cong Yu, Yang Hu, and Yan Chen. Rpm 2.0: Rf-based pose machines for multi-person 3d pose estimation. IEEE Transactions on Circuits and Systems for Video Technology, 34(1):490–503, 2023. 3 [63] Lili Xie, Hongchun Li, Jun Tian, Qian Zhao, Masahiro Shiraishi, Kenta Ide, Takahiro Yoshioka, and Takeshi Konno. Mikey: Human key-points detection using millimeter wave radar. In 2024 IEEE Wireless Communications and Networking Conference (WCNC), pages 1–6. IEEE, 2024. 3 [64] Qian Xie, Xinyu Hou, Qianyi Deng, Amir Patel, Niki Trigoni, and Andrew Markham. mmdiffusion: mmwave diffusion for sequential 3d human dense point cloud generation. In 2025 International Conference on 3D Vision (3DV), pages 781–790. IEEE, 2025. 3 [65] Hongfei Xue, Yan Ju, Chenglin Miao, Yijiang Wang, Shiyang Wang, Aidong Zhang, and Lu Su. mmmesh: Towards 3d real-time dynamic human mesh construction using millimeter-wave. In Proceedings of the 19th Annual International Conference on Mobile Systems, Applications, and Services, pages 269–282, 2021. 1, 3, 7 [66] Hongfei Xue, Qiming Cao, Yan Ju, Haochen Hu, Haoyu Wang, Aidong Zhang, and Lu Su. M4esh: mmwave-based 3d human mesh construction for multiple subjects. In Proceedings of the 20th ACM Conference on Embedded Networked Sensor Systems, pages 391–406, 2022. 3 [67] Hongfei Xue, Qiming Cao, Chenglin Miao, Yan Ju, Haochen Hu, Aidong Zhang, and Lu Su. Towards generalized mmwave-based human pose estimation through signal augmentation. In Proceedings of the 29th Annual International Conference on Mobile Computing and Networking, pages 1– 15, 2023. 3 [68] Kangwei Yan, Fei Wang, Bo Qian, Han Ding, Jinsong Han, and Xing Wei. Person-in-wifi 3d: End-to-end multiperson 3d pose estimation with wi-fi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 969–978, 2024. 2, 3, 7, 15 [69] Jianfei Yang, Yunjiao Zhou, He Huang, Han Zou, and Lihua Xie. Metafi: Device-free pose estimation via commodity wifi for metaverse avatar simulation. In 2022 IEEE 8th World Forum on Internet of Things (WF-IoT), pages 1–6. IEEE, 2022. 3 [70] Jiarui Yang, Songpengcheng Xia, Yifan Song, Qi Wu, and Ling Pei. mmbat: A multi-task framework for mmwavebased human body reconstruction and translation prediction. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 8446–8450. IEEE, 2024. 3 [71] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4210–4220, 2023. 2

[72] Dongqiangzi Ye, Yufei Xie, Weijia Chen, Zixiang Zhou, and Hassan Foroosh. Lpformer: Lidar pose estimation transformer with multi-task network. 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 16432–16438, 2023. 2 [73] Cunyi Yin, Xiren Miao, Jing Chen, Hao Jiang, Jianfei Yang, Yunjiao Zhou, Min Wu, and Zhenghua Chen. Powerskel: A device-free framework using csi signal for human skeleton estimation in power station. IEEE Internet of Things Journal, 11(11):20165–20177, 2024. 3 [74] Duo Zhang, Zhehui Yin, Xusheng Zhang, Junzhe Wang, Hongliu Yang, Zhiyun Yao, Zizhou Fan, Wenwei Li, and Daqing Zhang. Mupose: Breaking the scalability barrier of mmwave multi-user pose estimation in the wild. In Proceedings of the 24th Annual International Conference on Mobile Systems, Applications and Services, page 622–636, New York, NY, USA, 2026. Association for Computing Machinery. 18 [75] Jingyi Zhang, Qihong Mao, Guosheng Hu, Siqi Shen, and Cheng Wang. Neighborhood-enhanced 3d human pose estimation with monocular lidar in long-range outdoor scenes. Proceedings of the AAAI Conference on Artificial Intelligence, 38(7):7169–7177, 2024. 2 [76] Mingmin Zhao, Yonglong Tian, Hang Zhao, Mohammad Abu Alsheikh, Tianhong Li, Rumen Hristov, Zachary Kabelac, Dina Katabi, and Antonio Torralba. Rf-based 3d skeletons. In Proceedings of the 2018 Conference of the ACM Special Interest Group on Data Communication, pages 267–281, 2018. 2, 18 [77] Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11656–11665, 2021. 2 [78] Ce Zheng, Wenhan Wu, Chen Chen, Taojiannan Yang, Sijie Zhu, Ju Shen, Nasser Kehtarnavaz, and Mubarak Shah. Deep learning-based human pose estimation: A survey. ACM computing surveys, 56(1):1–37, 2023. 1 [79] Hongwei Zheng, Han Li, Wenrui Dai, Ziyang Zheng, Chenglin Li, Junni Zou, and Hongkai Xiong. Hipart: Hierarchical pose autoregressive transformer for occluded 3d human pose estimation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 16807–16817, 2025. 2 [80] Jingxiao Zheng, Xin Yu Shi, Alexander N. Gorban, Junhua Mao, Yang Song, C. Qi, Ting Liu, Visesh Chari, Andre Cornman, Yin Zhou, Congcong Li, and Drago Anguelov. Multi-modal 3d human pose estimation with 2d weak supervision in autonomous driving. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 4477–4486, 2022. 2 [81] Zhijie Zheng, Jun Pan, Zhikang Ni, Cheng Shi, Diankun Zhang, Xiaojun Liu, and Guangyou Fang. Recovering human pose and shape from through-the-wall radar images. IEEE Transactions on Geoscience and Remote Sensing, 60: 1–15, 2022. 2, 3 [82] Yue Zhou, Aichun Zhu, Caojie Xu, Fangqiang Hu, and Yifeng Li. Perunet: Deep signal channel attention in unet

for wifi-based human pose estimation. IEEE Sensors Journal, 22(20):19750–19760, 2022. 3 [83] Yunjiao Zhou, He Huang, Shenghai Yuan, Han Zou, Lihua Xie, and Jianfei Yang. Metafi++: Wifi-enabled transformerbased human pose estimation for metaverse avatar simulation. IEEE Internet of Things Journal, 10(16):14128–14136, 2023. 3 [84] Yunjiao Zhou, Jianfei Yang, He Huang, and Lihua Xie. Adapose: Toward cross-site device-free human pose estimation with commodity wifi. IEEE Internet of Things Journal, 11 (24):40255–40267, 2024. 3 [85] Bing Zhu, Zixin He, Weiyi Xiong, Guanhua Ding, Tao Huang, and Wei Xiang. Probradarm3f: mmwave radarbased human skeletal pose estimation with probability map guided multi-format feature fusion. IEEE Transactions on Aerospace and Electronic Systems, 2025. 3

RadPose: ISAC-based RF 3D Human Pose Estimation Supplementary Material In these appendices, we provide additional details on the monostatic ISAC system, the radio point cloud generation, and the network implementation. We also present additional experimental results under both normal and challenging scenarios, along with the failure cases of RadPose. Furthermore, we offer additional ablation studies of RadPose and discussions on its extension to multi-person pose estimation.

Appendix A: Radio Point Cloud Generation Beam sweeping

… Beam Beam pair 1 pair 2

Beam sweeping

Data transmission

Data Beam pair B

Data

Data

Data transmission

… Beam Beam pair 1 pair 2

Beam pair B

Figure 6. Beam-sweeping protocol in 5G NR systems.

Beam Sweeping in mmWave Communication. Both 5G FR2 and mmWave Wi-Fi systems periodically perform beam sweeping to establish and maintain directional communication links, as illustrated for 5G NR in Figure 6. During this procedure, the transmitter and receiver sequentially scan predefined codebooks of beam directions by transmitting and measuring reference or training signals. The receiver evaluates and record the signal from each candidate Tx-Rx beam pair, typically using metrics such as received signal strength or signal-to-noise ratio, and selects the pair that provides the best communication link. Meanwhile, the receiver estimates the channel associated with each beam pair. We denote the channel vector for beam pair b by h̃b , where 1 ≤ b ≤ B. Because user movement and environmental changes can alter the propagation channel, beam sweeping is performed periodically to maintain the quality of an active link. By leveraging the beam-sweeping protocol, we obtain the per-beam channel matrix H̃ = [h̃1 , h̃2 , . . . , h̃B ], where B is the total number of Tx–Rx beam pairs. For pointcloud generation, this matrix must be converted into the per-antenna channel matrix H = [hi,j ], where i and j index the Tx and Rx antenna elements, respectively. Because the beamforming weights associated with each beam pair are fixed, the relationship between the per-beam and per-antenna channel matrices can be calibrated in advance. Specifically, we recover the per-antenna channel matrix as H = TH̃, where T is a transformation matrix obtained through calibration. The resulting per-antenna channel matrix is then used to generate the radio point cloud, as de-

Image

S>1

S>2

Human

Human

Human

Human

Figure 7. Two samples of radio point clouds generated by our mmWave monostatic ISAC platform.

scribed in Section 3.2. Point Cloud Illustration. Eq. (3) presents the mathematical representation of a radio point cloud frame. Visualizing all points in a frame produces an overly dense point cloud, making the structure difficult to recognize. For clearer illustration, we define S>a = {|s| : s ∈ S, |s| > a} to display only the strong points in a radio frame, where S denotes the set of all points in the frame and a represents the point magnitude threshold. Figure 7 shows two examples of radio point cloud frames generated by our mmWave monostatic ISAC platform. The strong points visible in the images originate from both human targets and background objects. Our observations indicate that humans are strong reflectors of mmWave signals, generating prominent points in the radio frame. In addition to the static images, a demo video is provided on the project website, showing the radio point cloud in real time in dynamic environments.

Figure 8. Near-far effect in voxelization.

Figure 9. Illustration of physics-aware anisotropic sparse convolution.

Appendix B: Details of RadPose’s PhysicalAware Encoder The point cloud data acquired by an RF sensing system is naturally represented in spherical coordinates. In the spherical coordinate system as illustrated in Figure 8, the spatial extent covered by a voxel increases with radial distance

(i.e., the near-far effect). Conventional 3D convolution operations treat the voxel dimensions as isotropic, which is effective for feature extraction in Cartesian space but ineffective in spherical space. To address this issue, RadPose employs sparse convolution across three parallel branches (i.e., radial branch, angular branch, and point branch) for voxel-based feature extraction, as illustrated in Figure 9. The radial branch captures local continuity along the range axis, while the angular branch captures contextual patterns on the azimuthelevation plane. Furthermore, RadPose incorporates a point-wise residual branch to preserve local voxel responses without neighborhood mixing, thereby stabilizing feature fusion and reducing the risk of over-smoothing. The features encoded by the three branches are then linearly aggregated, followed by positional encoding in both spatial and temporal dimensions. The resulting encoded tokens are subsequently processed by spatial and temporal attention modules for further feature extraction.

Appendix C: Details of RadPose’s Heatmap Head Given the large number of voxels, the pose estimation network appears to have sparse gradients as the loss function is hard to capture every dimension of the input data. To address this issue, we adopt a 2.5D pose heatmap in the (r, θ) and (r, ϕ) planes as the output of the model, following the principle presented in [25], which provides denser information that can significantly reduce the gradient sparsity. To generate the 2.5D pose heatmaps, the conv/deconv head reshapes the target-frame tokens Xout into a spatial map and upsamples it using L transposed convolutions to produce the range-azimuth heatmap RA and elevation map E : F0 = Reshape(Xout ),

Fl+1 = σ(Deconv(Fl )),

(7)

RA = ConvRA 1×1 (FL ),

E = ConvE 1×1 (FL ),

(8)

where σ is an activation function and l = 0, 1, . . . , L − 1. For joint j in a human pose, its predicted coordinate pj is derived by soft-argmax on heatmap RAj and Ej . Let Ω be the set of its discrete locations on heatmap RA; denote RAj (v) as the heatmap value of its bin index v ∈ Ω. Then, the confidence of joint j is computed by:

cj = 1 −

! X 1 Pj (v) log (Pj (v)) , − log |Ω|

(9)

u∈Ω

RA (v)

j where Pj (v ∈ Ω) = P ′ RA , and ϵ is a small ′ j (v )+ϵ v ∈Ω value to avoid zero-division.

Appendix D: Details of RadPose’s Network Implementation RadPose voxelizes each input frame into 20,480 voxels, and five consecutive radio frames are fed into the network for pose estimation. Three layers of physics-aware sparse convolution are employed to encode each voxel into a 512-dimensional token. An attention-based downsampling layer then reduces the number of tokens from 20,480 to 2,048. Six layers of four-head Transformers are applied to model spatial dependencies. The spatial tokens from the five frames are separated into target-frame tokens and nontarget-frame tokens. For each non-target frame, the tokens are further downsampled to 576 tokens and concatenated with the target-frame tokens. Temporal dependencies are captured using two layers of four-head Transformers. The heatmap head consists of three convolutional layers followed by three deconvolutional layers to decode the joint heatmaps. For IK implementation, we use the 6th joint as the root joint and a three-layer GCN with 512 hidden dimensions as the IK head. Recall that RadPose has two key components in its network: pose estimator and GCN-based IK. When training the pose estimator, the loss weights are set to 100 for Lmap , 5 for Lxyz , 0.3 for Lang , and 3.5 for Llen . When training the GCN for IK, the loss weights are adjusted to 0.2 for Lang , 0.5 for Llen , and 3 for Lxyz . The weight of root loss Lroot is set to 0.5.

Figure 10. Illustration of RadPose performing sensing and communication simultaneously.

Appendix E: Data Collection and Overall System Setup RadPose is capable of performing sensing and communication simultaneously. Figure 1 illustrates how OFDM signals are utilized for both sensing and communication at the same time. During data collection, RadPose also transmits data packets to another device. Since the communication data path remains unchanged, the communication functionality and performance are preserved. Figure 10 shows an example where RadPose sends data packets to a receiver while performing sensing. In this experiment, a video was streamed from RadPose to the receiver. Table 7 summarizes the communication and sensing

Monostatic ISAC 120° 6m

(a) Set-up of RadPose

(b) Set-up of Wi-Fi in [68]

(c) Set-up of MVDoppler [10]

Figure 11. Experimental setups comparison of RadPose, Wi-Fi and mmWave radar.

parameters. The carrier frequency is 60 GHz. The bandwidth is 1.2288 GHz and the sampling rate is 1.2288 samples per second. The OFDM modulation uses 1024 subcarriers. Among the 1024 subcarriers, 900 are used for data transmission. The transmission power is 20 dBm. The field of view is 120◦ for horizon and 60◦ for elevation. Figure 11 shows the system and experimental setup of RadPose during our data collection. The figure also illustrates the experimental setup of our two baselines as a comparison. It is noteworthy that RadPose needs only one communication device for data collection, while the baselines need two or more communication devices for data collection. Table 7. The parameters of our monostatic ISAC device. Hardware parameters Sampling rate 1.2288 GSPS Number of Tx antennas 16 Number of Rx antennas 16 Center frequency 60 GHz Transmission power 20 dBm

Detection time/frame Theoretical range Effective ant. (hori.) Hori. antenna spacing Horizontal FoV

Communication parameters Waveform OFDM FFT points 1024 Valid subcarriers 900 Cyclic prefix length 276 OFDM symbol duration 1.057 µs OFDM symbols/frame 16 × 16 = 256 Supporting protocols 5G and Wi-Fi Sensing parameters 67 µs Frames per second 10 30 m Practical range 10 m 8 Effective ant. (elev.) 4 0.5λ Elev. antenna spacing 0.5λ [−60◦ , 60◦ ] Elevation FoV [−30◦ , 30◦ ]

Figure 12. Qualitative performance of RadPose in unseen scenes. Green pose represents ground-truth annotations from depth camera and blue one represents RadPose’s estimation results.

Appendix F: Qualitative Illustration of Estimated Poses In this section, we provide additional qualitative examples of poses estimated by RadPose to visually demonstrate the estimation accuracy and quality. Figure 12 presents the visualization results, covering various human activities such as walking, sitting, and more. The results demonstrate that RadPose accurately reconstructs skeletons that closely match the ground-truth from a depth camera. All examples are from previously unseen cases, highlighting the generalization capability of RadPose. A demo video is provided on the project website.

Appendix G: Qualitative Illustration of Robustness Studies Compared to vision-based sensors such as cameras and LiDAR, a key advantage of RadPose is its robustness to lighting conditions and occlusions. In this section, we present additional qualitative results highlighting these advantages. Figure 13 shows our experimental setup. In one scenario, RadPose was completely covered with materials such as cloth and foam, while the depth camera remained unobstructed. In another scenario, RadPose operated in a poorly lit room where the depth camera failed to produce human skeleton.

(a) Not cluded

Oc-

(b) Cloth Occluded

(c) Foam Occluded

(d) Dim Lighting

Figure 13. Experimental setup for different occlusion and lighting conditions.

Complete Blockage. Figure 14 presents several cases where the radio device antennas are completely covered by cloth and foam. In the figures, Green poses represent ground-truth annotations and blue poses represent estimated results. The results show that RadPose can still accurately estimate the pose of the subjects, thanks to the penetration capability of mmWave.

Figure 15. Qualitative performance of RadPose under low-lighting conditions where depth camera does not work. The brightness and contrast of RGB images have been adjusted to make it visible.

images, where joints become closely clustered. As a result, RF signals are prone to scattering-center merging and selfocclusion, which are difficult for a single-view sensor to resolve. Failures also occur when humans are close to background objects such as wall corners or metal railings. These objects have large radar cross-sections, causing strong reflections that mask the human body signals. Moreover, multipath effects exacerbate false detections and increase localization errors, particularly when subjects are near metallic surfaces.

(a) Cloth Occluded.

Figure 16. Illustration of RadPose’s failure cases in unseen scenes. (b) Foam Occluded.

Figure 14. Qualitative performance of RadPose occluded by cloth and foam.

Poor-Lighting Condition. Figure 15 presents the results under poor-lighting conditions. For illustration purpose, we manually enhanced he brightness and contrast of the RGB images to make the scenes visible. These adjusted RGB images serve as a reference for qualitative comparison. In this scenario, the depth camera fails to operate. In contrast, RadPose is capable of estimating the pose skeleton of the subjects. This is because RadPose does not rely on lighting illumination. We note that, since depth camera fails, there is no ground-truth annotations in the figure.

Appendix H: Qualitative Illustration of Failure Cases In this section, we present some failure cases of poses estimated by RadPose. Figure 16 illustrates several such instances. Most failure cases occur during human activities such as sitting and squatting, as shown in the first and third

Appendix I: Additional Experimental Results In this section, we present additional experimental results to analyze the performance of RadPose. Effectiveness of Heatmap Head. Instead of performing direct regression, RadPose employs a heatmap head to predict a 2.5D pose representation for supervised training. To evaluate the effectiveness of the heatmap head, we plot the training loss over iterations. Figure 17 presents the experimental results. The results show that, when using direct regression, the training of RadPose fails to converge due to sparse and unstable gradients. In contrast, when the proposed 2.5D heatmap-based loss is used, the training converges fast, demonstrating the effectiveness of the heatmap head. Ablation Study of PASC. RadPose employs PhysicsAware Sparse Convolution (PASC) as its encoder to mitigate the near-far effect. We conducted an ablation study to evaluate its effectiveness. We perform ablation studies to evaluate its effectiveness. Table 8 presents the results. Without PASC, RadPose achieves an MPJPE of

Table 10. Comparison of different voxels.

Figure 17. Training convergence comparison between direct regression and heatmap head. Table 8. Effectiveness of PASC. PASC

MPJPE (cm)↓

PCK10 ↑

w/o w

17.1 16.2

29.9 36.3

Table 9. Effectiveness of target-driven temporal attention (TDTA). Temporal Attention Mechanism No Temporal Attention Standard Temporal Attention Temporal Attention with Target Frame Query

MPJPE (cm)↓ PCK10 ↑ 20.8 22.0 16.2

6.1 7.5 36.3

17.1 cm, whereas incorporating PASC reduces the MPJPE to 16.2 cm. Although the gain is small, PASC is justified because it does not increase computational cost compared to conventional 3D convolution operations. Ablation Study of TDTA. Determined by the spectrum availability, the ranging resolution of RadPose is 12.2 cm, posing challenges in capturing reflections from small joints like wrists and ankles. Therefore, temporal attention becomes essential to improve the pose estimation accuracy. The target-driven temporal attention (TDTA) module in our model was designed to address this challenge. We now perform ablation studies to evaluate its effectiveness. Table 9 presents the pose estimation accuracy of RadPose using different temporal attention mechanisms. It includes two baselines: (i) no temporal attention and (ii) standard temporal attention for information aggregation. The results show that standard temporal attention does not improve performance and sometimes even degrades it, likely due to the introduction of additional noise from multiple frames. In contrast, RadPose’s TDTA approach (temporal attention with a target-frame query) yields a significant improvement for pose estimation accuracy. Impact of Voxelization. The voxel size introduces a tradeoff between accuracy and computational cost. On one hand, using a larger number of voxels can potentially improve performance; on the other hand, it significantly increases GPU memory consumption. Table 10 reports the measured MPJPE and the required GPU memory when RadPose uses different voxel configurations. Increasing the number of voxels from 20,480 to 30,720 improves MPJPE by only

Voxels

Memory (GB)

MPJPE(cm)↓

10240 20480 30720

4.9 6.4 7.9

19.7 16.2 15.9

0.3 cm. However, this configuration requires approximately 1.5 GB of additional GPU memory per input sample. Since this marginal accuracy gain does not justify the increased memory usage, we adopt 20,480 voxels as a balanced configuration. Impact of Frame Number. RF-derived environmental features are inherently sparse. To mitigate this sparsity, RadPose fuses multiple consecutive frames for pose estimation. Intuitively, incorporating more frames may improve performance by providing richer temporal information. To evaluate the impact of the number of frames, we conducted experiments by increasing the fusion window from three to seven frames. Table 11 presents the results. When five frames are used for pose estimation, RadPose achieves an MPJPE of 16.2 cm. However, increasing the number of frames to seven does not further improve accuracy; instead, the MPJPE slightly increases to 16.9 cm. This result suggests that incorporating excessive temporal information may introduce noisy or redundant features that degrade pose estimation performance. Table 11. Comparison of different frame numbers. Frames 3 5 7

Memory (GB)

MPJPE (cm)↓

4.5 6.4 8.3

18.6 16.2 16.9

Impact of Spatial Transformer Layers. After voxelization and encoding, RadPose uses multiple layers of spatial transformers to capture spatial relationships. Using more transformer layers can enhance representation capacity, but it also increases GPU memory consumption. To evaluate this trade-off, we conducted experiments by increasing the number of transformer layers from four to eight. Table 12 reports the results. Adding more layers is hard to further improve MPJPE, but it significantly increases GPU memory usage. Considering the trade-off between performance and computational cost, we adopt six layers in the final model configuration. Impact of Transformer Dimension. In addition to the number of transformer layers, the embedding dimension of the transformer also affects the performance of RadPose. To evaluate its impact, we conducted experiments by increasing the embedding dimension from 128 to 512. Table 13 reports our results. Increasing the dimension from 128 to 256 significantly improves pose estimation accuracy, reducing

Figure 18. A possible conceptual framework that extends RadPose to multi-person pose estimation.

Table 12. Comparison of different layers of transformer. Layers 4 6 8

Memory (GB)

MPJPE (cm)↓

5.8 6.4 6.9

21.4 16.2 16.1

the MPJPE from 23.6 cm to 18.0 cm. Further increasing the dimension to 512 yields the best performance, achieving an MPJPE of 16.2 cm. We adopt an embedding dimension of 512 in the final model to balance performance and computational cost. Table 13. Comparison of different dimensions of transformer. Dimensions 128 256 512

Memory (GB)

MPJPE (cm)↓

2.8 4.0 6.4

23.6 18.0 16.2

Appendix J: Computational Cost and Latency RadPose contains 174.9 million parameters and predicts one pose skeleton from a five-frame input window. We measure latency on preloaded data using CUDA synchronization, excluding storage I/O. In streaming inference, each incoming frame is processed using a rolling five-frame window, resulting in an end-to-end processing latency of 68.1 ms on a single GPU. Because radio data are acquired at 10 Hz, corresponding to a 100 ms sampling interval, this latency supports real-time operation with computation offloaded to an edge GPU. During inference, RadPose uses 6.46 GB of GPU memory, indicating that RadPose can be deployed on commonly available GPU accelerators without the substantially higher memory demands of training. Table 14 summarizes the model size, computational cost, resource usage, and latency.

Table 14. Computation cost and latency of RadPose inference. Property Parameters # of input frames Inference GPU Streaming latency Inference memory

Value 174.9M 5 Nvidia RTX 6000 Blackwell 68.1 ms 6.46 GB

Appendix K: Extending RadPose to MultiPerson Pose Estimation Currently, RadPose is trained and evaluated in single-person scenarios. Its framework could be extended to multiperson settings by, for example, following the approaches in [74, 76]. Specifically, a human-detection module could first identify person-specific regions within the dense radio point clouds and then apply RadPose’s single-person pose estimator to each detected region. Because RadPose reconstructs 3D poses from spatially structured radio point clouds without relying on fixed subject identities, the same pose estimation model can be shared across all detected regions. Several challenges must be addressed to realize this extension. RF responses from multiple subjects may overlap, while multipath propagation makes it difficult to associate each subject’s responses consistently across frames. Simply cropping individual regions and estimating their poses may discard useful contextual information, resulting in spatially misaligned or temporally inconsistent poses. One potential solution, illustrated in Figure 18, is to aggregate human-region proposals and associate them across time using learnable modules. Robustly addressing these challenges is beyond the scope of this work; therefore, we leave the implementation and evaluation of multi-person pose estimation to future work.

Related documents

• Radar-Based 3D Human Pose Estimation via Motion-Aware Pose Initialization and Observation-Guided Refinement#1403569
• A Two-Stage Shape- and Motion-Aware Framework for mmWave-based Human Mesh Recovery#1488364
• Doppler-aided Position-Velocity Tracking of Hidden Objects using mmWave Radar#1456908
• mmHRI: Towards Privacy-Preserving Human-Robot Interaction with Millimeter-Wave Radar#1119293
• mmIR: Frequency-Space Inverse Rendering for 3D Millimeter-Wave Radar ADC Synthesis#793146