betaFeedback

WorksEngineering

5G Sensing for Human Pose Estimation with PRS-based CSI: An Experimental Demonstration

Abstract

This paper presents an experimental study of device- free 2D human pose estimation using Channel State Information (CSI) extracted from downlink Fifth Generation (5G) Positioning Reference Signal (PRS). The proposed system employs a Next Generation Node Base (gNB) transmitting PRS to a User Equipment (UE) equipped with two receive antennas to measure the radio channel. The CSI is estimated using a least-squares (LS) method, and processed through amplitude smoothing and phase calibration. The resulting amplitude and phase representations are used as inputs to a deep learning model for pose estimation. We compare amplitude-only input with joint amplitude-phase training at a bandwidth of 20 MHz. The joint training outperforms the amplitude-only solution and achieves Percentage of Correct Keypoints (PCK) equal to 32.2%, 72.9%, and 84.2% below 10%, 30%, and 50% of the distance between left shoulder and right hip, respectively. These results demonstrate that standardized downlink 5G PRS-based CSI contains useful information for human pose estimation, even with a simple radio configuration and a relatively small dataset.

Introduction

Fifth Generation (5G) cellular systems are progressively evolving beyond their traditional role as communication networks to support positioning, sensing, and environmental awareness [1]. Standardized signals and positioning procedures provide a foundation for such capabilities [2], and recent research has highlighted the potential of leveraging physical-layer measurements of 5G networks for positioning [3] and sensing applications [4, 5, 6]. These developments align with the Integrated Sensing and Communication (ISAC) paradigm, where the same wireless infrastructure supports connectivity, localization, and environmental perception.

Among 5G signals, the Positioning Reference Signals (PRSs) play a central role in enabling network-based localization. They are transmitted by the Next Generation Node Bases (gNBs) and allow the receiver to estimate the gNB-to-User Equipment (UE) channel; their transmissions are specifically designed to facilitate time-of-arrival estimates through configurable resource allocation and flexible bandwidth configurations [7]. As a result, PRS measurements contain detailed information about the propagation environment and the objects interacting with the radio channel. While these signals have been extensively studied for positioning purposes, their suitability for fine-grained human sensing remains largely unexplored.

Recent advances in wireless sensing have demonstrated that radio-frequency channel measurements can capture information about human motion and activities [8, 9, 10]. In particular, deep learning approaches have shown that Channel State Information (CSI) can be directly mapped to human skeletal representations, enabling device-free pose estimation and activity recognition without requiring cameras or wearable devices [11, 12, 13, 14, 15, 16]. Most existing approaches, however, have been developed using Wi-Fi sensing platforms and datasets specifically designed for wireless perception. Consequently, their applicability to cellular systems operating under standardized 5G signaling remains an open research question.

The transition from Wi-Fi-based sensing to 5G sensing is not straightforward. In 5G, CSI is extracted from structured pilot resources within a network-controlled frame structure. The pilot allocation, bandwidth, transmitter geometry, and synchronization conditions are different from typical Wi-Fi sensing systems. Understanding whether CSI extracted from standardized 5G PRSs contains sufficient information for recovering human body pose is therefore an important step toward network-native sensing services. MiDiPose [6] represents a recent effort in this direction, using uplink Sounding Reference Signal (SRS)-based 5G measurements for 2D human pose estimation. It extracts amplitude, phase, and Doppler domain features from CSI, aligns these engineered feature vectors in a shared latent space, and fuses them for keypoint regression. However, downlink PRS-based CSI for 2D human pose estimation is still not well studied.

Instead of measurements collected from a multi-receiver configuration, we consider a setup consisting of a single-antenna gNB and a two-antenna UE. Although this configuration provides less spatial diversity, it allows us to investigate whether a single standardized PRS downlink already carries sufficient information for pose discrimination. While MiDiPose uses more than 200,000 samples in our case, we evaluate the proposed approach with a much lower sample number, to see whether accurate pose estimation can still be achieved with a short sensing time as typically required for activities performed in daily life.

In this work, we investigate human pose estimation from CSI extracted from downlink 5G PRS. A bistatic experimental platform based on Software-Defined Radios (SDRs) is employed to collect CSI measurements, to be used as input to a Deep Learning (DL) pose-estimation model originally developed for Wi-Fi sensing. The achieved results corroborate the experimental feasibility of recognizing human skeleton with 5G signals from standardized downlink PRS-based CSI in a configuration with a single gNB and two receive antennas, unlocking a new application area of the cellular wireless technology.

5g Sensing System and Dataset

Radio Sensing Scenario

We consider a 5G radio sensing system for human pose estimation, where PRSs are transmitted in downlink by one gNB and received by a UE equipped with two antennas. The scenario includes a human performing different poses in the sensing area. The received signals are first converted into CSI features, with amplitude and phase branches later used by the pose-estimation algorithm. At the same time, a camera supervises the scene, and outputs human skeleton keypoints using the YOLO26 algorithm [17].

The transmitter gNB is scheduled to transmit a PRS within one slot of a 5G frame. The PRS configuration uses a comb spacing of KCOMBPRS=2K_{\mathrm{COMB}}^{\mathrm{PRS}}=2 and LPRS=3L_{\mathrm{PRS}}=3, where LPRSL_{\mathrm{PRS}} denotes the number of Orthogonal Frequency-Division Multiplexing (OFDM) symbols carrying PRS within the scheduled slot. This pattern is illustrated in Figure 1, where the second slot is shown as an example. In the slot, three OFDM symbols are allocated for PRS pilots, represented by the green resource elements. The PRS pilots are modulated and transmitted over the air. The grey resource elements are not used for PRS in the current setup. They can be reserved for future ISAC extensions, for example by carrying 5G data symbols, or by supporting a future multistatic configuration with additional transmitting gNBs.

Representation of 5G frame, PRS comb-2 symbol allocation, and CSI extraction

Figure 1. Representation of 5G frame (20 slots at numerology 1), PRS comb-2 symbol allocation in a slot, and CSI extraction.

For each demodulated OFDM pilot symbol, the received resource-grid sample at UE receive antenna a∈{1,2}a \in\{1,2\}, OFDM-symbol index l∈{1,…,14}l \in\{1,\ldots,14\}, and subcarrier k=1,…,NKk=1,\ldots,N_K is denoted by Ya[k,l]Y_a[k,l]. Numerology μ=1\mu=1 is employed, corresponding to a Subcarrier Spacing (SCS) of 30 kHz, and 20 slots in a 5G frame of 10 ms [2]. The PRS bandwidth of B=20B=20 MHz is considered. Its expression in terms of Resource Blocks (RBs) is: NPRB=51N_{\mathrm{PRB}}=51 Physical Resource Blocks (PRBs). Therefore, the number of usable subcarriers is NK=12×NPRB=612N_K=12\times N_{\mathrm{PRB}}=612. The received samples are used for a least-squares (LS) channel estimation, and averaged during pre-processing per slot, then per frame.

CSI Estimation and Preprocessing

The received pilot sample at antenna aa, symbol ll and subcarrier kk is

Ya[k,l]=XPRS[k,l]Ha[k,l]+Wa[k,l],(1)Y_a[k,l] = X^{\mathrm{PRS}}[k,l]H_a[k,l] + W_a[k,l], \tag*{(1)}

where XPRS[k,l]X^{\mathrm{PRS}}[k,l] is the known transmitted PRS symbol, Ha[k,l]H_a[k,l] is the downlink channel frequency response from the gNB to receive antenna aa, and Wa[k,l]W_a[k,l] denotes noise and residual interference.

The symbol-level LS estimate of the PRS-based CSI is obtained independently for each receive antenna as

H^a[k,l]=Ya[k,l]XPRS[k,l].(2)\widehat{H}_a[k,l] = \frac{Y_a[k,l]}{X^{\mathrm{PRS}}[k,l]}. \tag*{(2)}

Since the PRS comb pattern only occupies a subset of subcarriers, the symbol-level estimates are first interpolated along the subcarrier dimension to obtain H^aint[k,l]\widehat{H}^{\mathrm{int}}_a[k,l], for k=1,…,NKk=1,\ldots,N_K.

In our processing, the CSI estimates are also organized at the slot level. For each slot ss, the interpolated PRS-based symbol-level estimates H^aint[k,l]\widehat{H}^{\mathrm{int}}_a[k,l] are averaged over the LPRSL_{\mathrm{PRS}} PRS OFDM symbols to obtain the slot-level estimate H^a[k,s]\widehat{H}_a[k,s].

CSI Preprocessing Steps

The estimated CSI comprises both amplitude and phase information that can be directly provided to the DL model. In addition, we evaluate two commonly adopted pre-processing operations: temporal Savitzky–Golay (SG) filtering of the amplitude and Linear Phase Removal (LPR) [16].

As concerns the amplitude, a polynomial order of two and a window length equal to approximately 0.2 times the input-vector length were found to provide the best performance. Following the same empirical principle, we treat 400 consecutive slot-level CSI estimates as one sequence and use a second-order polynomial with a window length of 81 samples.

The phase branch requires additional processing because the measured CSI phase is not only determined by the physical propagation channel [16]. It can also contain phase errors caused by hardware and synchronization effects. Previous CSI phase studies have shown that raw phase can contain random phase offsets and linear phase distortions, and that a LPR can be used to calibrate the measured phase [18]. In addition, a sampling timing offset introduces a phase term that is linear over the subcarrier domain [19]. Therefore, the measured phase is modeled as

Φa[k,s]=Φach[k,s]+θa[s]+γa[s]k+ξa[k,s],(3)\Phi_a[k,s] = \Phi_a^{\mathrm{ch}}[k,s] + \theta_a[s] + \gamma_a[s]k + \xi_a[k,s], \tag*{(3)}

where Φach[k,s]\Phi^{\mathrm{ch}}_{a}[k,s] is the channel-related phase, θa[s]\theta_{a}[s] an unknown common phase offset, γa[s]k\gamma_{a}[s]k the linear phase distortion across subcarriers, and ξa[k,s]\xi_{a}[k,s] a residual random phase noise.

The aim of the preprocessing phase is to reduce the non-channel phase terms in (3). For each slot ss, the unwrapped phase is fitted by the linear error component Φ^alin[k,s]=θ^a[s]+γ^a[s]k\widehat{\Phi}^{\mathrm{lin}}_{a}[k,s]=\widehat{\theta}_{a}[s]+\widehat{\gamma}_{a}[s]k, where the estimated offset and slope are obtained by

(θ^a[s],γ^a[s])=arg⁡min⁡θ,γ∑k=1NK(Φa[k,s]−θ−γk)2.(4)\left(\widehat{\theta}_{a}[s],\widehat{\gamma}_{a}[s]\right)=\underset{\theta,\gamma}{\arg\min}\sum_{k=1}^{N_K}\left(\Phi_a[k,s]-\theta-\gamma k\right)^2 . \tag*{(4)}

The fitted linear component can be removed by subtraction:

Φ~a[k,s]=Φa[k,s]−Φ^alin[k,s].(5)\widetilde{\Phi}_a[k,s]=\Phi_a[k,s]-\widehat{\Phi}^{\mathrm{lin}}_a[k,s]. \tag*{(5)}

When LPR is enabled, the phase representation is obtained as Φ~a[k,s]≈Φach[k,s]+ξa[k,s]\widetilde{\Phi}_a[k,s]\approx\Phi^{\mathrm{ch}}_a[k,s]+\xi_a[k,s], and the phase-corrected CSI is reconstructed as

H~a[k,s]=∣H^a[k,s]∣ejΦ~a[k,s].(6)\widetilde{H}_a[k,s]=\left|\widehat{H}_a[k,s]\right|e^{j\widetilde{\Phi}_a[k,s]}. \tag*{(6)}

Using the pre-processed CSI, the matrix is then constructed by concatenating along the subcarrier domain and slot domain with

H~at=[H~a[1,1]⋯H~a[1,Nslott]⋮⋱⋮H~a[NK,1]⋯H~a[NK,Nslott]]∈CNK×Nslott,(7)\widetilde{\mathbf{H}}_a^t= \begin{bmatrix} \widetilde{H}_a[1,1] & \cdots& \widetilde{H}_a[1,N_{\mathrm{slot}}^t]\\ \vdots& \ddots& \vdots\\ \widetilde{H}_a[N_K,1] & \cdots& \widetilde{H}_a[N_K,N_{\mathrm{slot}}^t] \end{bmatrix}\in\mathbb{C}^{N_K\times N_{\mathrm{slot}}^t}, \tag*{(7)}

with NslottN_{\mathrm{slot}}^t denoting number of slots in one CSI sample tt. After concatenating also the antenna domain aa, the CSI tensor used as the learning input has size H~t∈C2×NK×Nslott\widetilde{\mathbf{H}}^t\in\mathbb{C}^{2\times N_K\times N_{\mathrm{slot}}^t}. The amplitude and the unwrapped phase along the subcarriers are obtained from the CSI tensor as [∣H~t∣ ∠H~t]\left[\left|\widetilde{\mathbf{H}}^t\right| \ \angle\widetilde{\mathbf{H}}^t\right]. These two tensors are then used as selectable input branches for the DL model estimating the pose.

Pose Prediction Models and Pipeline

The full learning pipeline from PRS-based CSI extraction to 2D pose prediction is illustrated in Figure 2, while the experimental acquisition and CSI-to-camera alignment are detailed in Section III-C. On the radio side, the two-antenna UE estimates the CSI from the received PRS, followed by preprocessing to obtain the amplitude and phase inputs, while the camera provides the corresponding 2D skeleton as ground truth. The amplitude ∣H~t∣\left|\widetilde{\mathbf{H}}^t\right| and phase ∠H~t\angle\widetilde{\mathbf{H}}^t branches are independently encoded by the Masked Autoencoder (MAE) encoders, then fused and passed through the task-prompting block, Graph Convolutional Network (GCN), and Transformer modules to predict the final human pose.

Overview of the proposed DT-Pose framework

Figure 2. Overview of the proposed DT-Pose framework for 5G CSI based 2D human pose estimation using two USRPs with camera-based GT supervision. The 2-antenna receiver is highlighted on the left, the single antenna transmitter is behind the camera.

DL Model

DT-Pose [14] is adapted as a 2D pose-estimation backbone while keeping its main design unchanged. Following the original two-stage idea, we first pretrain the MAE-Vision-Transformer (ViT) encoder with a patch size of (NK/NPRB,4)(N_K/N_{\mathrm{PRB}},4) in a self-supervised manner. Since the proposed input has two possible branches, the amplitude and phase branches are pretrained separately. After pretraining, the MAE encoders are used to extract branch-specific representations. For joint amplitude-phase training, the two branches are then combined by a learnable fusion:

Mt=α1Z~ampt+α2Z~phaset,(8)\mathbf{M}^t=\alpha_1\widetilde{\mathbf{Z}}_{\mathrm{amp}}^t+\alpha_2\widetilde{\mathbf{Z}}_{\mathrm{phase}}^t, \tag*{(8)}

with α1+α2=1\alpha_1+\alpha_2=1, Z~ampt\widetilde{\mathbf{Z}}_{\mathrm{amp}}^t denoting the amplitude branch after MAE processing and Z~phaset\widetilde{\mathbf{Z}}_{\mathrm{phase}}^t the phase after MAE. For amplitude-only training, Mt=Z~ampt\mathbf{M}^t=\widetilde{\mathbf{Z}}_{\mathrm{amp}}^t. Therefore, the model input keeps the common form Mt∈R2×NK×Ns\mathbf{M}^t\in\mathbb{R}^{2\times N_K\times N_s}.

The downstream DT-Pose decoder follows the original topology-aware design. It uses task prompts, GCN, and transformer layers to model the structure of the human skeleton. The task prompt provides a learnable pose-aware query for the decoder. The GCN layers then inject the human skeleton topology into the decoding process. The transformer layers capture long-range dependencies among human joints [14].

The learning target is a 2D human skeleton, described according to the Common Objects in Context 17 (COCO17) format. Accordingly, the graph-convolution part exploits the COCO17 bone topology, adjacent joints and explicitly modeling the spatial relationship between connected body limbs considered during pose decoding. The final regression layer is modified to output 2D keypoints, where each output corresponds to the 2D coordinates of a specific anatomical landmark of the human body, such as the nose, shoulders, elbows, wrists, hips, knees, and ankles.

Loss Function

For each camera frame, the model predicts 2D human skeleton keypoints p^∈RJ×2\widehat{\mathbf{p}}\in\mathbb{R}^{J\times2}, where J=17J=17 is the number of keypoints. The camera image resolution is 640×480640\times480. After extracting the 2D skeleton pixels, they are scaled by a factor to represent a meter scale, assuming the human is about 170 cm and is moving on a plane at a fixed camera depth. The corresponding Ground Truth (GT) pose is denoted by p∈RJ×2\mathbf{p} \in\mathbb{R}^{J \times2}. The coordinate loss is the mean squared error of the keypoint estimate averaged over all frames:

System parameterSymbol / keyValue
Numerologyμ\mu1
BandwidthBB20 MHz
Number of PRBsNPRBN_{\mathrm{PRB}}51
Number of subcarriersNKN_{\mathrm{K}}612
Subcarrier spacingΔf\Delta f30 kHz
PRS combKCOMBPRSK_{\mathrm{COMB}}^{\mathrm{PRS}}2
PRS symbolsLPRSL_{\mathrm{PRS}}3
Number Rx antennasNRxN_{\mathrm{Rx}}2
Camera frame ratefcamf_{\mathrm{cam}}8.66 Hz
5G frames per sampleNfN_f2
5G slots per sampleNslottN_{\mathrm{slot}}^t40
Number of keypointsJJ17
Sensing area–5.3×4.3 m25.3 \times4.3\ \mathrm{m}^2

Table I. SYSTEM PARAMETERS OF THE 5G PRS-BASED CSI DATASET

LMSE=1NtJ∑t=1Nt∑j=1J∥p^t,j−pt,j∥22.(9)\mathcal{L}_{\mathrm{MSE}} = \frac{1}{N_tJ}\sum_{t=1}^{N_t}\sum_{j=1}^{J}\left\lVert\hat{\mathbf{p}}_{t,j}-\mathbf{p}_{t,j}\right\rVert_2^2. \tag*{(9)}

Here, the camera frame number NtN_t also denotes the number of training frames. All models are trained using the AdamW optimizer with a learning rate of 10−410^{-4} and weight decay rate of 10−210^{-2}. The batch size is chosen as 64 and the model is trained for 200 epochs. The dataset is divided into training, validation, and test subsets using non-overlapping temporal blocks with an approximate ratio of 7:2:17 : 2 : 1 so that adjacent samples from same activity segments are not shared across subsets. The final model is selected based on the lowest validation loss.

Validation Metrics

The primary evaluation metrics are validation loss and Percentage of Correct Keypoints (PCK). The PCK measures the average keypoint localization accuracy over all frames and all joints. It is defined as

PCK@τ=1NtestJ∑t=1Ntest∑j=1J1(∥p^t,j−pt,j∥2∥pt,ls−pt,rh∥2≤τ).(10)\mathrm{PCK}@\tau= \frac{1}{N_{\mathrm{test}}J}\sum_{t=1}^{N_{\mathrm{test}}}\sum_{j=1}^{J}\mathbf{1}\left(\frac{\left\lVert\hat{\mathbf{p}}_{t,j}-\mathbf{p}_{t,j}\right\rVert_2}{\left\lVert\mathbf{p}_{t,\mathrm{ls}}-\mathbf{p}_{t,\mathrm{rh}}\right\rVert_2}\leq\tau\right). \tag*{(10)}

where NtestN_{\mathrm{test}} is the number of test frames, JJ is the number of keypoints, and p^t,j\hat{\mathbf{p}}_{t,j} and pt,j\mathbf{p}_{t,j} are the predicted and GT 2D coordinates of keypoint jj in frame tt, respectively. The function 1(⋅)\mathbf{1}(\cdot) is a binary indicator. The normalization term ∥pt,ls−pt,rh∥2\left\lVert\mathbf{p}_{t,\mathrm{ls}}-\mathbf{p}_{t,\mathrm{rh}}\right\rVert_2 is the distance between the left shoulder and the right hip [13], [14]. In this work, we compute the PCK at τ∈{10%,30%,50%}\tau\in\{10\%, 30\%, 50\%\}, meaning the percentage of keypoints with errors below the corresponding percentage of the reference torso distance.

Model Evaluation

Experimental Data Collection

The measurements were collected in an indoor sensing area of 5.3×4.3 m25.3 \times4.3\ \mathrm{m}^2, with the parameters for the 5G and camera systems as reported in Table I and the USRPs share the same 10 MHz reference synchronization signal provided by an Ettus OctoClock-G CDA 2290. As illustrated in Figure 2, one USRP acts as gNB with one transmit antenna and another USRP acts as UE equipped with two receive antennas. The yellow circle indicates the placement of the receive antennas. The transmit antenna was placed on the opposite edge of the sensing area, such that the human was located between the transmitter and receiver. A single participant started approximately at the midpoint between the two devices and performed a continuous sequence of eight everyday activities. A set of 8 different activities is considered, numbered as indicated in Figure 3 and in Table III, with each activity repeated at least three times in each time interval. The human moved to the left side and returned to the initial position, then did repetitions of shaking hands, bending the body, stretching legs, rotation, crouching, jumping and falling.

Preprocessed CSI over the full activity sequence

Figure 3. Preprocessed CSI over the full activity sequence: (a) amplitude and (b) phase after linear component removal. The dashed vertical lines indicate the activity segments. The numbers indicate from 1 to 8: moving, shaking, bending, stretching, rotating, crouching, jumping, falling.

The camera provides pose labels at fcam=8.66f_{\mathrm{cam}} = 8.66 Hz with one camera frame every 115 ms. The keypoints are linearly interpolated by a factor of 4 between camera frames to obtain a frame of 28.9 ms. For numerology μ=1\mu= 1, one 5G radio frame has duration 10 ms. During the LS estimation, several frames are lost because some synchronization blocks are not recognized, reducing the 5G radio samples by a factor of 0.7. Therefore, one camera frame corresponds to about Nslott=2×20=40N_{\mathrm{slot}}^t = 2 \times20 = 40 slots. These slots are aligned with the camera frame, then taken into one CSI sample tt. With this setup, two experiments are recorded, denoted as β1\beta_1 or β2\beta_2. Experiment β1\beta_1 represents the primary measurement campaign with continuous activities, while experiment β2\beta_2 is recorded with faster activities and the recording is not continuous.

To illustrate the characteristics of the processed CSI, Figure 3 shows the channel response amplitude and phase over frequency and over the full activity sequence in experiment β1\beta_1. As seen in the figure, both components exhibit clear temporal variations across different activities. The horizontal axis represents time in seconds, while the vertical axis represents the subcarrier index. The dashed lines mark the transitions between consecutive activities, and the corresponding activity labels are shown at the top of each plot. In Figure 3a, the amplitude presents activity-dependent temporal variations across the subcarriers. Although the overall amplitude remains relatively smooth, clear vertical patterns appear during several activity intervals. In particular, the longer falling interval produces a visibly different response from shorter activities such as shaking, rotating, and jumping. Figure 3b shows the phase after preprocessing and temporal smoothing. The processed phase also contains activity-dependent temporal structures, although a small number of samples still show residual phase outliers close to −180∘-180^\circ or 180∘180^\circ.

Using these preprocessed inputs, we evaluate DT-Pose for the two independent 20 MHz recordings and compare the amplitude-only and joint amplitude–phase input representations. We then analyze the results by activity and keypoint, examine the training behaviour, and finally present representative qualitative pose estimates.

Training and Validation Performance

The learning behaviour of the joint amplitude–phase model is shown in Figure 4. Over the epochs, the training and validation losses gradually decrease, while the validation PCK values improve and eventually become stable. The curves show no clear late-stage divergence between training and validation performance, suggesting that the model continues to learn useful pose-related features. Most of the improvement occurs before approximately epoch 180. After training, the metrics are listed in Table II together with the number of available samples after interpolation.

Training and validation loss and validation PCK curves

Figure 4. Training and validation performance in terms of loss and PCK for joint amplitude–phase training. The validation loss is used to select the final model checkpoint.

Exp.PreprocessingSamplesPCK@10 ↑PCK@30 ↑PCK@50 ↑
β1\beta_1None12 67751.5 / 44.183.2 / 80.891.2 / 90.1
β1\beta_1LPR12 67741.0 / 42.476.3 / 79.686.2 / 88.8
β1\beta_1SG12 67744.5 / 39.177.5 / 78.787.4 / 88.5
β1\beta_1SG+LPR12 67750.0 / 48.283.2 / 83.190.7 / 90.7
β2\beta_2SG+LPR407230.8 / 19.869.5 / 54.374.6 / 70.0

Table II. Comparison of PCK under different CSI preprocessing configurations and experimental recordings. Each entry is reported as joint amplitude–phase training / amplitude-only training.

For the experiment β1\beta_1, the results show that LPR does not consistently improve pose-estimation performance. Without SG filtering, retaining the linear phase component improves the joint amplitude–phase PCK from 41.0%, 76.3%, and 86.2% to 51.5%, 83.2%, and 91.2% at PCK@10, PCK@30, and PCK@50, respectively. This suggests that the fitted linear phase component may contain useful propagation-related information in addition to synchronization-induced distortions. In contrast, when SG filtering is applied, LPR improves PCK@10 from 44.5% to 50.0%, PCK@30 from 77.5% to 83.2% and PCK@50 from 87.4% to 90.7%, indicating that an interaction between temporal smoothing and phase detrending might help. These results highlight that the SG filter and LPR preprocessing together do not provide any benefit compared to DL applied directly to raw amplitude and phase data. However, combining SG filter and LPR improves the joint amplitude–phase performance compared with either operation used individually. This indicates that the combined SG+LPR configuration remains beneficial relative to the single-operation alternatives. Finally, for the β2\beta_2 experiment, joint amplitude–phase training consistently outperforms amplitude-only training despite the smaller dataset.

Pose Results

Table III reports the per-activity performance of the 20 MHz joint amplitude–phase model. The model performs best for bending and moving, while rotating, jumping, and falling are more difficult. These activities involve larger body motion, rapid transitions, or more complex limb configurations, which can increase keypoint localization errors. Furthermore, ten representative pose-estimation examples are illustrated in Figure 5, highlighting examples of successful (Figure 5a) and imperfect (Figure 5b) estimates. For each example, the first row shows the camera image with the GT skeleton, while the second row shows the corresponding pose predicted from the PRS-based CSI. The successful examples in Figure 5a show that the proposed pipeline can recover a range of body configurations, including standing, bending, crouching, falling, and arm-raising poses. In most cases, the predicted skeleton follows the main body orientation and preserves the relative positions of the torso and limbs. This indicates that the CSI input contains useful spatial information about human motion and posture.

NumberActivityPCK@10 ↑PCK@30 ↑PCK@50 ↑
1Moving84.1597.2899.56
2Shaking60.2487.3393.44
3Bending68.4893.0597.24
4Stretching59.3588.3894.30
5Rotating37.2372.7682.11
6Crouching50.1384.3793.26
7Jumping30.1969.7284.30
8Falling30.3072.9285.48

Table III. Per-activity performance assessed as in (10) for the 20 MHz joint amplitude–phase model. PCK values are reported in percentage, and higher values are better.

Representative qualitative pose estimation results

Figure 5. Representative qualitative pose estimation results. For each example, the first row shows the camera image with the GT skeleton, while the second row shows the pose predicted by the CSI-based pipeline. The red regions highlight joints with large prediction errors.

The failure cases in Figure 5b mainly occur at limb endpoints, especially the wrists and ankles. In the first and second failure cases, there is one joint having a poor alignment to the GT. For example, in the third case, the subject is rotating, but the predicted skeleton faces the wrong direction. For this type of activity, which includes rotation, the depth dimension plays a significant role. Additional spatial diversity, for example through multiple gNBs or receive antennas, may help resolve such orientation ambiguities. A similar orientation error appears in the fourth crouching example. Overall, the model estimates the main body structure well, while fine-grained limb localization and orientation remain challenging in the 2D domain with a small number of dataset samples.

Conclusion

This work investigated whether CSI extracted from standardized 5G PRS can support device-free 2D human pose estimation. At 20 MHz, joint amplitude–phase training improves strict keypoint localization compared with amplitude-only training, with the largest gain observed for PCK@10. The pre-processing comparison further showed that dedicated SG filter and LPR are not required with the best performance obtained by directly using the amplitude and phase of the LS-estimated CSI. The proposed pipeline also recovers the main body structure across different activities. The wrists and elbows are more difficult to localize because they often move rapidly and cover a larger spatial range. Some failure cases also contain errors in body orientation or hand positions. Future work will investigate generalization across multiple subjects, environments, activity sequences, and radio geometries, using larger and different bandwidths. More robust phase processing, longer training schedules and static pose recordings may further improve the localization of rapidly moving joints. Additional gNBs or receive antennas could also provide greater spatial diversity.

Acknowledgment

The authors acknowledge the use of an AI-assisted tool (ChatGPT) for English-language proofreading. All AI-generated content was carefully reviewed and revised by the authors to ensure scientific accuracy and consistency.

References

  1. [1]E. S. Lohan, “Emerging topics in joint radio-based positioning, sensing, and communications,” Sensors, vol. 25, no. 3, p. 948, 2025.DOI
  2. [2]3GPP, “NR; Physical channels and modulation,” 3rd Generation Partnership Project (3GPP), Technical Specification TS 38.211, 2020.
  3. [3]M. Brambilla et al., “Integration of 5G and GNSS technologies for enhanced positioning: An experimental study,” IEEE Open Journal of the Communications Society, vol. 5, pp. 7197–7215, 2024.DOI
  4. [4]E. Özbay, P. K. Bishoyi, and M. Petrova, “Empowering 5G PRS-based ISAC with compressed sensing,” in 2024 IEEE 25th International Workshop on Signal Processing Advances in Wireless Communications (SPAWC), 2024, pp. 341–345.arxiv.org/abs/2407.13478
  5. [5]Y. E. Sadgudu, K. Davasliglu, T. Erpek, S. Kompella, G. Anderson, and J. Ashdown, “MULTI-SCOUT: Multistatic integrated sensing and communications in 5G and beyond for moving target detection, positioning, and tracking,” in MILCOM 2025-2025 IEEE Military Communications Conference (MILCOM), 2025, pp. 693–698.
  6. [6]H. Li, D. Li, A. Zhang, W. Zhang, C. Sun, and H. Zhang, “No vision, no wearables: 5G-based 2D human pose recognition with integrated sensing and communications,” arXiv preprint arXiv:2512.24923, 2025.arxiv.org/abs/2512.24923
  7. [7]L. Italiano, B. Camajori Tedeschini, M. Brambilla, H. Huang, M. Nicoli, and H. Wymeersch, “A tutorial on 5G positioning,” IEEE Communications Surveys & Tutorials, vol. 27, no. 3, pp. 1488–1535, 2025.
  8. [8]S. Savazzi, M. Nicoli, F. Carminati, and M. Riva, “A bayesian approach to device-free localization: Modeling and experimental assessment,” IEEE Journal of Selected Topics in Signal Processing, vol. 8, no. 1, pp. 16–29, 2013.DOI
  9. [9]S. Savazzi, S. Sigg, M. Nicoli, V. Rampa, S. Kianoush, and U. Spagnolini, “Device-free radio vision for assisted living: Leveraging wireless channel quality information for human sensing,” IEEE Signal Processing Magazine, vol. 33, no. 2, pp. 45–58, 2016.DOI
  10. [10]I. Guarino, D. Carra, M. Cominelli, F. Gringoli, and R. L. Cigno, “A survey on CSI-based Wi-Fi sensing datasets and models with a focus on reproducibility,” Computer Communications, p. 108431, 2026.DOI
  11. [11]F. Wang, S. Zhou, S. Panev, J. Han, and D. Huang, “Person-in-WiFi: Fine-grained person perception using WiFi,” in 2019 IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 5451–5460.
  12. [12]J. Geng, D. Huang, and F. De la Torre, “Densepose from WiFi,” arXiv preprint arXiv:2301.00250, 2022.arxiv.org/abs/2301.00250
  13. [13]Y. Zhou, H. Huang, S. Yuan, H. Zou, L. Xie, and J. Yang, “MetaFi++: WiFi-enabled transformer-based human pose estimation for metaverse avatar simulation,” IEEE Internet of Things Journal, vol. 10, no. 16, pp. 14 128–14 136, 2023.DOI
  14. [14]Y. Chen, J. Guo, S. Guo, J. Zhou, and D. Tao, “Towards robust and realistic human pose estimation via WiFi signals,” arXiv preprint arXiv:2501.09411, 2025.arxiv.org/abs/2501.09411
  15. [15]Z. Yin, M. Brambilla, and M. Nicoli, “Human walking sensing and pose estimation in the 6 GHz band using amplitude and phase CSI,” arXiv preprint arXiv:2606.10048, 2026.arxiv.org/abs/2606.10048
  16. [16]G. Diaz, I. Sobron, I. Eizmendi, I. Landa, J. Coyote, and M. Velez, “Channel phase processing in wireless networks for human activity recognition,” Internet of Things, vol. 24, p. 100960, 2023.DOI
  17. [17]R. Sapkota, R. H. Cheppally, A. Sharda, and M. Karkee, “YOLO26: Key architectural enhancements and performance benchmarking for real-time object detection,” arXiv preprint arXiv:2509.25164, 2025.arxiv.org/abs/2509.25164
  18. [18]X. Wang, L. Gao, and S. Mao, “PhaseFi: Phase fingerprinting for indoor localization with a deep learning approach,” in 2015 IEEE Global Communications Conference (GLOBECOM), 2015, pp. 1–6.
  19. [19]M. Kotaru, K. Joshi, D. Bharadia, and S. Katti, “SpotFi: Decimeter level localization using WiFi,” ACM SIGCOMM Computer Communication Review, vol. 45, no. 4, pp. 269–282, 2015.

Paper details

vlnc:W_ganqqs7f
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 9

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

ChatGPT

View statement
Looking for related papers…
Contents