PoseGBDT: Temporal UWB Posture Classifier

• The paper demonstrates that augmenting LightGBM with sliding-window temporal features boosts accuracy from 81.58% to 99.11% in posture recognition.
• It fuses multi-modal UWB data including ranging, signal quality, and CIR features to capture both coarse and fine-grained posture cues.
• The approach incorporates an OOD detector using One-Class SVM on GBDT leaf embeddings to robustly handle dynamic transitions and environmental variations.

Searching arXiv for the PoseGBDT source paper and related context papers. {"command":["bash","-lc","python - <<'PY'\nimport urllib.request, urllib.parse\nquery='all:PoseGBDT OR ti:\"UWB-PostureGuard\"'\nurl='http://export.arxiv.org/api/query?search_query='+urllib.parse.quote(query)+'&start=0&max_results=5'\nprint(url)\ntry:\n print(urllib.request.urlopen(url, timeout=20).read().decode('utf-8')[:4000])\nexcept Exception as e:\n print('ERR', e)\nPY"]} {"command":["bash","-lc","python - <<'PY'\nimport urllib.request, urllib.parse\nqueries=['all:PoseGBDT','ti:\"UWB-PostureGuard\"','id:(Li et al., 14 Aug 2025)','all:\"Online Gradient Boosting Decision Tree\"']\nfor query in queries:\n url='https://export.arxiv.org/api/query?search_query='+urllib.parse.quote(query)+'&start=0&max_results=3'\n print('\nQUERY',query)\n try:\n data=urllib.request.urlopen(url, timeout=20).read().decode('utf-8')\n print(data[:1800])\n except Exception as e:\n print('ERR', e)\nPY"],"timeout_ms":120000} {"command":["bash","-lc","curl -L --max-time 20 'https://export.arxiv.org/api/query?search_query=all:PoseGBDT&start=0&max_results=3' | sed -n '1,80p'"],"timeout_ms":120000} {"command":["bash","-lc","python - <<'PY'\nimport requests\nfor url in ['https://export.arxiv.org/api/query?search_query=all:PoseGBDT&start=0&max_results=1','https://arxiv.org/abs/([2508.11115](/papers/2508.11115))']:\n try:\n r=requests.get(url, timeout=20)\n print(url, r.status_code, r.text[:300])\n except Exception as e:\n print('ERR', url, e)\nPY"],"timeout_ms":120000} {"command":["bash","-lc","echo hello"]} PoseGBDT is a temporal-aware classifier introduced in the UWB-PostureGuard system for continuous, contactless monitoring of ergonomic sitting posture. It is built on the LightGBM implementation of Gradient Boosted Decision Trees (GBDT), but its defining feature is not a new tree-construction rule; rather, it is the use of temporally augmented ultra-wideband (UWB) radiofrequency features, including ranging information, channel impulse response (CIR), and signal quality metrics, together with an out-of-distribution (OOD) detection stage based on One-Class SVM over GBDT leaf embeddings. In the reported real-world evaluation across 10 participants and 19 distinct postures, the system achieved 99.11% accuracy while maintaining robustness against clothing thickness, additional devices, and furniture configurations (Li et al., 14 Aug 2025).

1. Position within UWB-PostureGuard

PoseGBDT was proposed as the core classifier in a privacy-preserving RF sensing pipeline for ergonomic sitting posture monitoring. The surrounding system uses commercial UWB devices and targets a setting in which camera-based methods raise privacy concerns and wearable sensors impose user discomfort. Within that system design, PoseGBDT addresses the classification problem by operating on rich RF-derived features and explicitly modeling short-term temporal dependencies that frame-wise classifiers omit (Li et al., 14 Aug 2025).

The method is motivated by a specific limitation of standard frame-wise classification. Treating each sample independently overlooks posture transitions and short-term context, which in turn limits discrimination between similar or dynamically shifting poses. PoseGBDT addresses this by constructing sliding-window inputs that combine current observations with historical features and rolling statistics. This design places the method in a hybrid regime: the predictor remains a conventional LightGBM GBDT, but the input representation is temporally structured.

A common misunderstanding is to treat PoseGBDT as merely “GBDT for posture.” More precisely, it is a temporal feature-construction strategy coupled to LightGBM for posture recognition. The empirical comparison against standard frame-wise LightGBM is therefore central to the method’s identity, because the reported improvement comes from temporal augmentation rather than from replacing GBDT with a different ensemble learner (Li et al., 14 Aug 2025).

2. RF feature representation and engineered feature space

PoseGBDT operates on a multi-modal UWB feature space. The feature engineering pipeline combines ranging features, signal quality features, and CIR-derived features. The ranging group includes Distance, Azimuth/Elevation, FOM, and PDoA. The signal quality group includes the NLoS indicator, First/Main Path Index, Signal-to-Noise Ratios, RSSI, CIR Main/First Path Power, Noise Variance, Carrier Frequency Offset, and AoA Phase. The CIR group includes complex CIR samples and derived magnitude and phase representations.

Feature group	Included attributes	Role
Ranging	Distance, Azimuth/Elevation, FOM, PDoA	Geometric and angle-related posture cues
Signal quality	NLoS indicator, path index, SNR, RSSI, power, noise variance, CFO, AoA phase	Channel-condition and propagation cues
CIR-based	Complex CIR, magnitude, phase	Fine-grained RF signature structure

The CIR representation is written as

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

From this complex sequence, PoseGBDT uses both magnitude $|h[n]|$ and phase $\angle h[n]$. Post-processing includes IQR-based denoising for robust discrimination (Li et al., 14 Aug 2025).

This feature design is notable because it does not rely on a single RF statistic. Instead, it fuses geometric measurements, path-quality indicators, and raw channel-response structure. A plausible implication is that the classifier’s high performance depends on complementary observables: ranging and AoA-type features encode coarse posture geometry, whereas CIR-derived terms capture finer variations induced by body position and occlusion patterns.

3. Temporal modeling with LightGBM GBDT

PoseGBDT models temporal dependency through sliding-window feature construction. For each frame at index $i$, the input includes the current feature vector $\mathbf{x}_i$, time-lagged features $\mathbf{x}_{i-k}$ for lags $k = 1 \ldots \tau-1$, and a rolling window mean. The input construction is given as

$$\left[ \mathbf{x}_i, \mathbf{x}_{i-1}, \mathbf{x}_{i-2}, \ldots, \mathbf{x}_{i-(\tau-1)}, \bar{\mathbf{x}}_i^{(w)} \right],$$

with

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

This temporalization enables the model to exploit short-term posture dynamics, smooth transient noise, and distinguish transitions that would be ambiguous under per-frame classification. In the reported ablation, the selected window size was $\tau = 5$ (Li et al., 14 Aug 2025).

The underlying boosting model is standard GBDT. After $|h[n]|$0 boosting steps, the model is

$|h[n]|$1

where $|h[n]|$2 is the number of leaves in tree $|h[n]|$3, $|h[n]|$4 is region $|h[n]|$5 in tree $|h[n]|$6, and $|h[n]|$7 is the output for that region. For multiclass posture recognition, optimization uses cross-entropy:

$|h[n]|$8

The important methodological point is that PoseGBDT does not alter the boosting objective itself. Temporal dependency is injected at the representation level, after which a standard LightGBM learner performs multiclass classification. This distinguishes the method from sequential neural architectures while still allowing temporal sensitivity.

4. Inference pipeline and OOD detection

The end-to-end PoseGBDT pipeline begins with raw UWB streams and proceeds through feature extraction, temporal aggregation, classification, and OOD filtering. The paper summarizes the algorithm in five steps: input of raw UWB data streams; feature extraction from ranging, CIR, and signal quality; sliding-window creation; LightGBM classifier training with cross-entropy loss; and OOD detection using One-Class SVM on LightGBM leaf embeddings (Li et al., 14 Aug 2025).

The OOD component is based on internal tree-ensemble structure rather than on raw RF features alone. For a test sample, the method collects its traversal path across all trees in the trained LightGBM forest, forming a leaf-node embedding. A One-Class SVM is then fit to the distribution of in-distribution leaf embeddings so that unmodeled transitions or postures can be detected.

This design clarifies another potential misconception. OOD detection in PoseGBDT is not an auxiliary confidence threshold on softmax-like outputs; it is a separate detector defined over GBDT-internal representations. Because the embedding is derived from the sample’s path through the forest, it reflects the ensemble’s partitioning of the temporally augmented feature space rather than only the final class score.

5. Experimental protocol, accuracy, and robustness

The reported evaluation used 10 volunteers, 19 postures, and 1710 samples corresponding to 171,000 frames. The train-test split was 60% training and 40% testing, stratified to preserve posture classes. Metrics included Overall Accuracy, Precision, Recall, F1-Score, and confusion matrices. The LightGBM hyperparameters were Number of leaves = 64, Learning rate = 0.05, Early stopping = 10 steps, and Loss = Cross-entropy, with $|h[n]|$9 selected via ablation (Li et al., 14 Aug 2025).

Method	Macro F1 (%)	Accuracy (%)
Standard frame-wise LightGBM	81.69	81.58
PoseGBDT (Temporal)	99.05	99.11

The performance gap between frame-wise LightGBM and PoseGBDT is the central empirical result. The frame-wise baseline achieved F1 = 81%, whereas PoseGBDT with temporal features and $\angle h[n]$0 achieved F1 = 99.05%, described as a greater than 18% absolute increase. The confusion matrices reportedly show a marked reduction in misclassification, especially for easily confused dynamic classes such as distinguishing left and right leg shakes (Li et al., 14 Aug 2025).

The reported per-regime behavior is also specific. Static postures achieved metrics above 99%, while dynamic postures exceeded 97% F1, compared with approximately 60–71% for the baseline. The system retained greater than 90% accuracy under environmental variations including thick clothing, electronic devices, furniture rearrangement, and placement variability. Under more severe obstructions or interference, such as a laptop in front, a smartphone on a table, or antennas being misaligned, accuracy dropped to 85–93%.

Continuous monitoring was demonstrated at a 0.2 s frame interval, and the OOD detector reportedly identified unknown or unlabeled postures rather than misclassifying them. These results support the interpretation that PoseGBDT is designed not only for closed-set posture recognition but also for deployment conditions involving transitions and unmodeled postures.

6. Conceptual significance and relation to adjacent GBDT work

PoseGBDT exemplifies a broader pattern in recent GBDT research: adapting tree ensembles to problem classes often associated with more specialized model families. Recent work has introduced in-place incremental and decremental learning for GBDT (Lin et al., 3 Feb 2025) and an efficient Hessian estimator that enables GBDTs in stochastic learning to rank (Kang et al., 2024). This suggests that PoseGBDT should be read as part of a wider expansion of GBDT methodology into dynamic and structured settings, although its own contribution is specifically temporal feature augmentation for posture recognition rather than online tree updates or new second-order estimation machinery.

At the same time, PoseGBDT should not be conflated with similarly named but unrelated methods. “Prototype Guided Backdoor Defense” addresses backdoor sanitization in deep neural networks through prototype geometry and post-hoc fine-tuning (Amula et al., 26 Mar 2025), whereas PoseGBDT concerns UWB-based posture classification with LightGBM and OOD detection (Li et al., 14 Aug 2025). The shared acronymal resemblance does not imply methodological overlap.

The central methodological claim of PoseGBDT is therefore narrower and more concrete: temporal modeling, achieved via augmented GBDT input, is critically important for real-world, contactless posture monitoring using multi-modal UWB features. Within the evidence presented, the method shows that carefully engineered temporal context can enable a conventional LightGBM classifier to achieve very high performance across static and dynamic postures in realistic environments, while preserving the privacy-preserving, camera-free, and wearable-free operating assumptions of the surrounding UWB-PostureGuard system (Li et al., 14 Aug 2025).