UWB-PostureGuard: Contactless Ergonomic Monitor

• The paper presents a contactless posture monitoring system that uses paired UWB devices and the PoseGBDT model to achieve 99.11% accuracy across 19 distinct sitting postures.
• UWB-PostureGuard is defined by a four-stage process, integrating ranging and CIR data with time-lagged features to robustly classify ergonomic sitting postures under privacy constraints.
• The design employs a feedback loop with OOD rejection and a sliding temporal window, ensuring dynamic posture changes are effectively detected and misclassifications minimized.

UWB-PostureGuard is a privacy-preserving, contactless sitting posture monitoring system built on commercial ultra-wideband (UWB) devices for continuous ergonomic sitting posture monitoring (Li et al., 14 Aug 2025). It targets the longstanding trade-off between observability and acceptability in posture assessment by using RF sensing rather than cameras or wearable instrumentation. In the reported formulation, paired UWB devices collect ranging and Channel Impulse Response (CIR) data from the user’s environment, a temporal-aware Gradient Boosting Decision Tree model termed PoseGBDT performs posture classification and out-of-distribution (OOD) detection, and a feedback loop notifies users about unhealthy postures while incorporating their reactions into ongoing posture assessment (Li et al., 14 Aug 2025).

1. System definition and architectural scope

UWB-PostureGuard is defined as a four-stage pipeline consisting of UWB Data Collection, Feature Extraction, PoseGBDT Model Inference, and a Feedback/Notification Loop (Li et al., 14 Aug 2025). The sensing stack uses 2× Murata Type2BP UWB modules (3×3 cm) and 1× Raspberry Pi 5 controller, with an Optional: Raspberry Pi Camera for ground truth in experimental setup (Li et al., 14 Aug 2025). In the standard deployment described for experiments, the paired UWB devices are placed on the edges of a desk, typically placed 1 m apart at a height of 1 m with antennas oriented toward the user (Li et al., 14 Aug 2025).

The stated design objective is continuous ergonomic sitting posture monitoring under privacy, comfort, and deployment constraints (Li et al., 14 Aug 2025). Privacy follows from the fact that the system processes RF reflections rather than images or audio; comfort follows from the absence of wearables; and scalability is tied to the use of small commercial modules and a compact desk-mounted configuration (Li et al., 14 Aug 2025). The article’s framing of UWB-PostureGuard as a mobile-health and preventive-health system is therefore grounded in both sensing modality and deployment geometry rather than in a generic human-activity-recognition abstraction.

A useful way to situate the system is against adjacent UWB posture-sensing paradigms. SleepPoseNet uses an off-the-shelf single antenna UWB radar for non-contact sleep postural transition recognition (Piriyajitakonkij et al., 2020), while wearable range-only posture recognition for human-robot interaction uses five body-mounted UWB nodes and all-to-all distances as input to KNN, SVM, and MLP classifiers (Salimi et al., 2024). UWB-PostureGuard differs from both by focusing on continuous ergonomic sitting posture monitoring, using a paired desk-edge UWB setup and a temporal-aware LightGBM-based classifier with OOD rejection rather than radar-image classification or purely wearable range-only classification (Li et al., 14 Aug 2025).

2. Sensing configuration and signal observables

The system gathers both ranging and CIR data from the desktop environment (Li et al., 14 Aug 2025). The ranging features are obtained through Double-Sided Two-Way Ranging (DS-TWR) and include:

• Distance between UWB initiator and anchor, $d$, in cm.
• Azimuth/Elevation angles $(\theta, \phi)$, described as angle-of-arrival estimates characterizing spatial displacement.
• FOM (Figure of Merit), described as a reliability metric for the angle estimate.
• PDoA, the phase difference of arrival in degrees from the antenna pair (Li et al., 14 Aug 2025).

The signal-quality features comprise SNR Main Path / First Path / Total, RSSI, and Path Indices for the first and main path time indices (Li et al., 14 Aug 2025). The CIR is treated as a complex-valued propagation descriptor,

$h[n] = \text{Re}(h[n]) + j \cdot \text{Im}(h[n]),$

from which both Magnitude $|h[n]|$ and Phase $\angle h[n]$ are derived (Li et al., 14 Aug 2025). Noise suppression is performed with Inter-quartile range (IQR) filtering (Li et al., 14 Aug 2025).

This sensing formulation is consistent with broader findings that posture- and pose-dependent RF effects are not reducible to raw distance alone. In the dataset study on relative pose impact in UWB localization, ranging bias is modeled as

$$\tilde{d} = d + b(\tilde{d}, \bm{\alpha}, \bm{\beta}, \bm{\gamma}) + n,$$

with empirical results showing that azimuth and elevation angles and the difference angle between tag and anchor have a dominant effect on bias, whereas range alone is not the major driver (Choe et al., 2024). That observation provides a direct contextual rationale for UWB-PostureGuard’s inclusion of angle-related, signal-quality, and CIR-derived features rather than relying on scalar distance only.

The choice to exploit CIR is also aligned with other UWB posture- and pose-sensing work. A purely UWB-based yaw-estimation method predicts orientation directly from anchor-side CIR using a bidirectional Mamba architecture and achieves 18.9 degrees mean absolute error with a location-based Kalman filter (Cheraghinia et al., 24 Jun 2026). Smartphone pose detection for a tagless gate similarly combines UWB CIR with inertial data and reports 0.984 LOS/NLOS classification accuracy and 0.961 four-pose detection accuracy in real time (Choi et al., 2024). These results suggest that the CIR is not merely an ancillary diagnostic but a high-value posture-conditioned observable within UWB-PostureGuard’s sensing regime.

3. Feature engineering and temporal representation

UWB-PostureGuard uses a deliberately heterogeneous feature space spanning ranging features, signal quality features, CIR-based features, and temporal enrichment (Li et al., 14 Aug 2025). The feature-engineering strategy is central because the target problem is not only posture classification in isolated frames but also the discrimination of fine-grained ergonomic states and dynamic transitions.

The temporal enrichment step adds time-lagged features and rolling statistics over a sliding window $(\tau=5$ frames) (Li et al., 14 Aug 2025). The explicit formulations are

$$\mathbf{x}_i^{(\text{shift}, k)} = \mathbf{x}_{i - k}, \quad \bar{\mathbf{x}_i^{(w)}} = \frac{1}{w} \sum_{j=0}^{w-1} \mathbf{x}_{i-j}.$$

The reported optimum window is $\tau=5$ frames, described as empirically balancing granularity and overfitting (Li et al., 14 Aug 2025). The final input vector includes features for the current frame, past 4 frames, rolling mean over past 4 frames (Li et al., 14 Aug 2025).

This design choice directly addresses a weakness of frame-wise posture inference. The paper states that PoseGBDT was developed to capture temporal dependencies in posture patterns, addressing limitations of traditional frame-wise classification approaches (Li et al., 14 Aug 2025). The importance of temporal modeling is corroborated by several related UWB systems. SleepPoseNet improves over a WRTFT-only deep CNN by jointly exploiting time-domain and frequency-domain representations for sleep postural transitions, achieving $73.7 \pm 0.8 \%$ mean accuracy versus $(\theta, \phi)$0 for the referenced baseline (Piriyajitakonkij et al., 2020). In wearable IMU+UWB human motion tracking, Group Inertial Poser uses a structured state-space model to learn temporal dynamics from acceleration, rotation, and intra-user UWB distances (Xue et al., 24 Oct 2025). Although these systems address different targets, they converge on the same methodological point: posture- and motion-related inference benefits materially from temporally structured representations.

A further implication is that UWB-PostureGuard should be understood as an RF time-series classification system rather than as a purely geometric range estimator. Its signal pathway uses desk-mounted UWB as a proxy for posture-conditioned changes in propagation geometry, angle estimates, and multipath structure (Li et al., 14 Aug 2025). This is closer in spirit to RF sensing than to classical anchor-tag localization, even though it reuses ranging primitives such as DS-TWR.

4. PoseGBDT, temporal classification, and OOD rejection

The core inference model is PoseGBDT, whose backbone is LightGBM (a Gradient Boosting Decision Tree implementation), selected for superior performance with high-dimensional, heterogeneous features and robustness to noise (Li et al., 14 Aug 2025). The boosting update is given as

$(\theta, \phi)$1

with training under Softmax cross-entropy

$(\theta, \phi)$2

The reported model configuration uses 64 leaves per LightGBM tree and learning rate 0.05 (Li et al., 14 Aug 2025).

The OOD component is a One-Class SVM on LightGBM leaf embeddings, used to learn a manifold for valid posture patterns and reject unfamiliar, transitional, or noisy postures (Li et al., 14 Aug 2025). In the system architecture, this is not a peripheral add-on but part of the continuous-use design, because real deployments involve transitions, incomplete postures, and off-taxonomy states that would otherwise be forced into an in-distribution class. The paper reports that PoseGBDT, with its OOD extension, reliably identifies unseen posture states/transitions, avoiding misclassifications during real-world continuous monitoring sequences (Li et al., 14 Aug 2025).

This combination of boosted tree inference with OOD rejection is methodologically distinct from the dominant neural-sequence alternatives in the surrounding literature. SleepPoseNet uses a multi-view CNN architecture over time- and frequency-domain radar features (Piriyajitakonkij et al., 2020). UWB CIR-based yaw detection uses a bidirectional Mamba model with Kalman filtering (Cheraghinia et al., 24 Jun 2026). Smartphone and tagless-gate pose detection use CNN or CNN-LSTM pipelines conditioned on LOS/NLOS state (Choi et al., 2024, Choi et al., 2024). Wearable IMU+UWB motion systems use structured state-space models, BiLSTMs, UKFs, or diffusion models (Xue et al., 24 Oct 2025, DeVrio et al., 3 Sep 2025, Liu et al., 14 May 2025, Hollidt et al., 1 Jun 2026). UWB-PostureGuard therefore occupies a comparatively unusual point in the design space: tabular temporal feature engineering plus boosted-tree classification plus explicit OOD rejection (Li et al., 14 Aug 2025).

A plausible implication is that this design favors interpretability and robustness under moderate dataset size, especially when the features already encode physics-relevant observables such as CIR magnitude/phase, PDoA, SNR structure, and sliding-window statistics. That implication is consistent with the reported preference for LightGBM under high-dimensional heterogeneous inputs, although the paper does not frame the result as an interpretability study (Li et al., 14 Aug 2025).

5. Dataset, posture taxonomy, and reported performance

The evaluation uses a standard office desk/chair/table arrangement with multiple environmental variations, including earbuds, smartphones, smartwatches, pillow, backpack, thick clothing, laptop, water bottle, alternative chair types, and multiple UWB device placements (distance, orientation, height) (Li et al., 14 Aug 2025). The posture taxonomy contains 19 distinct postures, described as including upright, lean forward/back, cross-legged, hunched, etc. (Li et al., 14 Aug 2025). The dataset contains 10 participants, 1,710 samples (171,000 frames), where each sample consists of 100 consecutive UWB frames for a specific posture and scenario (Li et al., 14 Aug 2025). Training and evaluation use a 60/40 train-test split, stratified by posture, with Accuracy, Precision, Recall, and F1-score (both macro and weighted) as the main metrics (Li et al., 14 Aug 2025).

The principal results are summarized below.

Quantity	Reported result	Source
Participants	10	(Li et al., 14 Aug 2025)
Postures	19 distinct postures	(Li et al., 14 Aug 2025)
Samples / frames	1,710 samples / 171,000 frames	(Li et al., 14 Aug 2025)
Overall accuracy	99.11%	(Li et al., 14 Aug 2025)
Weighted F1-score	99.11%	(Li et al., 14 Aug 2025)
Temporal window	$(\theta, \phi)$3 frames	(Li et al., 14 Aug 2025)

The paper further reports that all individual postures (including dynamic and similar postures) exceed 97% F1-score (Li et al., 14 Aug 2025). For difficult dynamic classes, it gives the example that "horizontal leg shake" improves from ~57%→98% relative to framewise LightGBM (Li et al., 14 Aug 2025). Baselines are listed as Naive Bayes, k-NN, Random Forest, and MLP (Li et al., 14 Aug 2025), but the supplied summary does not provide their full score table; it only states that PoseGBDT outperforms these comparators under the reported protocol.

These metrics place UWB-PostureGuard at the high-accuracy end of UWB-based posture-recognition results in the supplied literature, though direct cross-paper comparison remains imperfect because sensing geometry, posture taxonomy, and evaluation protocols differ. SleepPoseNet reports $(\theta, \phi)$4 mean accuracy for four sleep postural transitions using single-antenna UWB radar (Piriyajitakonkij et al., 2020). Wearable range-only posture recognition for nine postures reports 97% for SVM and 96% for MLP under leave-one-subject-out cross-validation with five on-body UWB nodes (Salimi et al., 2024). UWB-PostureGuard reports 99.11% overall accuracy for 19 distinct postures in a contactless desktop setup (Li et al., 14 Aug 2025). This suggests that the desk-mounted RF-sensing formulation, combined with feature engineering and temporal modeling, is highly effective for seated ergonomic monitoring within the evaluated office domain.

6. Robustness, deployment constraints, and relation to broader UWB posture sensing

Robustness is reported across several environmental perturbations. The paper states that earbuds, backpacks, water bottles: no significant decrease (~97%+ accuracy); smartphone in pocket/on desk or using a laptop: moderate drop (to ~90%/85% accuracy) due to RF interference or blocking; thick clothing: accuracy ~90%, confirming UWB’s penetration; and different device layouts: some accuracy loss (antenna height 91.35%), but performance robust across standard office arrangements (Li et al., 14 Aug 2025). These observations define the practical operating envelope more concretely than the aggregate accuracy alone.

The privacy and comfort claims are tightly bounded in the paper. Privacy is grounded in the fact that the system collects only RF echoes, preventing recovery of personal identity, appearance, or environment, and that no vision or audio is ever captured or processed (Li et al., 14 Aug 2025). Comfort is grounded in the claim that the method is completely contactless: No cameras, no device attachment, and works in all clothing/furniture conditions, with passive operation—no active participation required from the user (Li et al., 14 Aug 2025). Adaptability is supported by robustness across objects, body types, and device placements, and by the compact hardware footprint described as $(\theta, \phi)$5 (Li et al., 14 Aug 2025).

Within the broader UWB literature, UWB-PostureGuard occupies one branch of a larger posture- and pose-sensing continuum. At one extreme are RF-only, non-contact systems such as SleepPoseNet for sleep transitions (Piriyajitakonkij et al., 2020) and CIR-based orientation detection using bidirectional Mamba (Cheraghinia et al., 24 Jun 2026). At another are wearable range-only systems that use multiple on-body UWB nodes for posture recognition and robot control (Salimi et al., 2024). A third family fuses UWB with IMUs for full-body or arm-pose reconstruction, including Group Inertial Poser (Xue et al., 24 Oct 2025), UMotion (Liu et al., 14 May 2025), Ultra Diffusion Poser (Hollidt et al., 1 Jun 2026), and SmartPoser (DeVrio et al., 3 Sep 2025). UWB-PostureGuard is distinguished by four characteristics: contactless operation, desk-mounted commercial hardware, ergonomic sitting posture taxonomy, and PoseGBDT with OOD rejection (Li et al., 14 Aug 2025).

Several cross-paper findings sharpen the interpretation of its design. First, relative pose and orientation can systematically bias UWB ranging, which supports the inclusion of angle, PDoA, and CIR features rather than treating distance as sufficient (Choe et al., 2024). Second, temporal continuity improves RF posture inference, whether implemented through CNN fusion (Piriyajitakonkij et al., 2020), Kalman filtering (Cheraghinia et al., 24 Jun 2026), or LightGBM with lagged and rolling features (Li et al., 14 Aug 2025). Third, UWB posture sensing remains sensitive to blockage, device placement, and environmental structure, as reflected in both the desktop robustness results of UWB-PostureGuard and the LOS/NLOS-aware designs of tagless-gate pose systems (Choi et al., 2024, Choi et al., 2024).

A plausible implication is that future variants of UWB-PostureGuard could incorporate stronger geometric or sequence constraints, as seen in IMU+UWB motion systems that reconstruct body-consistent sensor layouts or propagate uncertainty online (Liu et al., 14 May 2025, Hollidt et al., 1 Jun 2026). The reported system, however, is specifically a continuous ergonomic sitting posture monitoring platform, and its principal contribution lies in showing that paired commercial UWB devices, comprehensive feature engineering, temporal boosted-tree inference, and OOD detection can achieve 99.11% accuracy across 10 participants and 19 distinct postures in a contactless office setting (Li et al., 14 Aug 2025).