PaperlayerView exact revision on arXiv

Source-linked AI summary

Structured Pose-Conditioned Flow Matching for Generative 5G CSI Augmentation

Haojin Li, Anbang Zhang, Wai Ho Mow, Chenyuan Feng, Chen Sun, Haijun Zhang

arXiv:2609.29912v1eess.SPcs.AI

TL;DR

Limited and costly synchronized CSI-pose data constrain 5G human pose recognition. StructFlow-HPR generates pose-conditioned CSI in a structured latent space, and the paper reports aligned generated samples and improved HPR performance with augmentation.

  • Problem

    Collecting and annotating large synchronized CSI-pose datasets is costly, while HPR models trained on few real pairs can overfit and generalize poorly.

    Source: §I · p. 1

  • Method

    StructFlow-HPR uses a topology-preserving autoencoder and pose-guided latent transport to generate CSI samples aligned with pose conditions.

    Source: §I · p. 1, A. Structured Latent Representation · p. 3

  • Results

    Average PCK rises from 72.43% to 73.71% at PCK@5 and from 91.48% to 92.84% at PCK@10 with StructFlow augmentation, while generated and real latent samples largely overlap across illustrated actions.

    Source: 1) Generation Quality Evaluation · p. 4, 3) Real-World Validation of StructFlow… · p. 5

  • Takeaways & Limitations

    The paper concludes that pose-aligned CSI augmentation improves downstream HPR under limited-data conditions and identifies adaptive action-specific, cross-environment, and multi-person extensions as future work.

    Source: §V · p. 5

Abstract

from arXiv · show

With the growing demand for privacy-preserving and occlusion-resilient human pose recognition (HPR), 5G channel state information (CSI) offers a promising contactless sensing modality by integrating communication and sensing capabilities. However, collecting large-scale synchronized CSI-pose pairs remains costly in practical 5G systems. To address this limitation, we propose StructFlow-HPR, a structured pose-conditioned flow matching framework for generative CSI augmentation. StructFlow-HPR learns a continuous latent transport process from Gaussian noise to real CSI representations under pose guidance, while preserving the receiver-frequency topology of CSI through a reconstruction-preserving autoencoder. A pose-conditioned Transformer is further designed to model the latent velocity field and generate pose-aligned CSI samples via ordinary differential equation sampling. Experiments on real-world 5G sensing data show that StructFlow-HPR can produce realistic CSI-pose pairs and improve downstream HPR performance under limited-data conditions.

I. INTRODUCTION

Optical sensing faces line-of-sight, illumination, and occlusion challenges, while collecting synchronized CSI-pose data is costly. StructFlow-HPR uses pose-guided generation while preserving CSI structure, and the paper describes CSI features and aligned pose labels.

I. INTRODUCTION

  • Optical sensing is limited by line-of-sight conditions, illumination changes, and physical occlusions, while synchronized high-precision CSI-pose data is costly to acquire and annotate.The paper presents generative augmentation as a response to data scarcity in HPR.Source: §I · p. 1
  • 5G CSI captures human-motion-induced signal variation through integrated sensing and communication, with spatial multiplexing and multipath features supporting fine-grained motion sensing.Source: §I · p. 1
  • StructFlow-HPR learns pose-guided latent transport from Gaussian noise to real CSI and generates pose-aligned samples while preserving CSI receiver-frequency structure.It combines a topology-aware autoencoder with a pose-window conditioned Transformer velocity network for downstream HPR augmentation.Source: §I · p. 1

A. 5G Collaborative Sensing and CSI-Pose Representation

  • Motion features combine windowed amplitude statistics, inter-receiver phase differences, and Doppler-related information from temporal phase variation.The phase-difference feature is motivated by its greater stability than absolute phase; the Doppler calculation uses the sampling interval.Source: A. 5G Collaborative Sensing and… · p. 2
  • Each CSI frame is organized as a tensor with subcarrier, receiver, and feature-channel dimensions, and paired with a synchronized pose label containing keypoints and coordinates.Source: A. 5G Collaborative Sensing and… · p. 2

B. Problem Formulation for Generative CSI Augmentation

With few real CSI-pose pairs, downstream HPR can overfit and generalize poorly. The formulation treats augmentation as pose-conditioned CSI generation, then combines synthetic and real data for training.

B. Problem Formulation for Generative CSI Augmentation

A. Structured Latent Representation

StructFlow-HPR uses a structured latent representation to address CSI dimensionality and receiver-frequency dependencies, then conditions generation on a window of nearby poses to capture local motion context.

A. Structured Latent Representation

  • A topology-preserving autoencoder encodes CSI into a compact latent representation because raw CSI is high-dimensional and has strong receiver-frequency dependencies.Source: A. Structured Latent Representation · p. 3
  • The structured latent space preserves CSI receiver-frequency topology, providing a physically meaningful manifold for generative modeling.Source: A. Structured Latent Representation · p. 3
  • For each target CSI frame, the pose-window condition uses neighboring pose labels rather than only the current pose, with the window vector flattened, normalized, and embedded by an MLP.This local context is intended to capture short-term motion continuity and provide richer semantics than a single pose snapshot.Source: A. Structured Latent Representation · p. 3

B. Pose-Conditioned Flow Matching and CSI Augmentation

StructFlow-HPR learns a pose-conditioned velocity field to transport Gaussian noise toward real CSI latents, then samples synthetic CSI-pose pairs by ODE integration for HPR augmentation.

B. Pose-Conditioned Flow Matching and CSI Augmentation

  • The Transformer estimates conditional latent transport dynamics, with the flow-matching objective training its velocity predictions against the target velocity.The model conditions on pose windows and flow time; it learns latent dynamics rather than predicting diffusion noise.Source: B. Pose-Conditioned Flow Matching and… · p. 3
  • Sampling initializes latent states from Gaussian noise, evolves them with the learned ODE, and decodes the result into CSI paired with its conditioning pose.Repeating this process creates the synthetic pairs used for augmentation.Source: B. Pose-Conditioned Flow Matching and… · p. 3
  • The downstream HPR model trains on a mixed dataset and is evaluated exclusively on held-out real CSI-pose samples.This evaluation checks generalization rather than memorization of synthetic patterns.Source: B. Pose-Conditioned Flow Matching and… · p. 3

1) 5G-enabled Prototype Platform:

The study uses a fixed indoor multi-node 5G sensing platform to collect synchronized CSI-pose pairs for daily activities. It defines real-only and real-plus-generated training settings and evaluates methods on the same held-out real test set.

1) 5G-enabled Prototype Platform:

  • The indoor platform places a transmitting smartphone and three receiving RRUs at room corners, covering a 3.3 m×2.7 m area with transceivers fixed at 1.5 m height.The base station uses H3C equipment, with Sony Xperia 1 IV smartphones as transmitting UEs.Source: 1) 5G-enabled Prototype Platform · p. 3

2) Data Collection and Augmentation Protocol:

  • The system synchronizes 5G uplink SRS-based CSI with visual pose sequences for five daily activities, aligning frames by nearest timestamp.The activities include squatting, hand raising, leg raising, leg pressing, and body translation.Source: 2) Data Collection and Augmentation… · p. 3
  • Table I compares MetaFi, MfDfHPR, and StructFlow-HPR across PCK thresholds and motion states.The table note maps A1, A2, and A3 to those three methods, respectively.Source: Table I · p. 4, 2) Data Collection and Augmentation… · p. 4
  • Fig. 2 visualizes real and StructFlow-generated CSI latent distributions for representative motion states using t-SNE.The caption identifies the visualization scope but does not state a comparison outcome.Source: Figure 2 · p. 4
  • Collecting raw full-scale data for one continuous action category takes approximately 30 minutes; comparisons use real-only or real-plus-generated training data and a shared held-out real test set.The real-only setting excludes the held-out segment, while the augmented setting adds generated samples to all available real training samples.Source: 2) Data Collection and Augmentation… · p. 4

3) Compared Methods and Evaluation Protocol:

StructFlow-HPR is evaluated against MetaFi and MfDfHPR under a shared backbone and representation dimensionality. Its generated latent samples largely overlap real samples across representative motions, while Fig. 3 compares pose coordinates from the visual model and StructFlow-HPR.

3) Compared Methods and Evaluation Protocol:

  • StructFlow-HPR adds a pose-conditioned flow-matching generator to the MfDfHPR framework while retaining the same downstream pose-regression architecture.All methods also use standardized representation dimensionality and the same MetaFi neural backbone to constrain computational complexity and memory usage.Source: 3) Compared Methods and Evaluation… · p. 4

1) Generation Quality Evaluation:

  • Generated samples largely overlap real CSI latents across Squat, Move, and Rise Hand, indicating pose-conditioned distribution learning with major CSI manifold structure and reasonable diversity preserved.The samples are assessed after encoding and t-SNE projection.Source: 1) Generation Quality Evaluation · p. 4
  • Fig. 3 presents human pose coordinates generated by the visual model and StructFlow-HPR, respectively.The caption identifies the compared sources but does not report which coordinates are more accurate.Source: Figure 3 · p. 5
  • Table I reports action-wise HPR performance at multiple PCK thresholds, with PCK@5 the most stringent setting and larger thresholds more relaxed.PCK@α measures the percentage of predicted keypoints within α% of torso length.Source: 2) Comparison With Baseline HPR Models · p. 5

2) Comparison With Baseline HPR Models:

StructFlow-generated CSI-pose augmentation improves average HPR PCK over MfDfHPR across four thresholds. The paper concludes that it improves downstream HPR under limited-data conditions and identifies adaptive action-specific and cross-environment and multi-person sensing as future directions.

2) Comparison With Baseline HPR Models:

  • MfDfHPR’s average PCK reaches 72.43%, 91.48%, 98.30%, and 99.08% at PCK@5, PCK@10, PCK@20, and PCK@30, respectively.Compared with MetaFi, MfDfHPR raises PCK@5 from 67.84% to 76.93% on Squat and from 34.51% to 43.36% on Press Leg2.Source: 2) Comparison With Baseline HPR Models · p. 5

3) Real-World Validation of StructFlow Augmentation:

  • Average PCK rises from 72.43% to 73.71% at PCK@5, 91.48% to 92.84% at PCK@10, 98.30% to 98.84% at PCK@20, and 99.08% to 99.38% at PCK@30.For Press Leg2, PCK@5, PCK@10, and PCK@20 improve by 4.50%, 5.13%, and 3.33%, respectively.Source: 3) Real-World Validation of StructFlow… · p. 5

V. CONCLUSIONS AND FUTURE WORK

  • The conclusion identifies adaptive action-specific augmentation and extension to cross-environment and multi-person 5G sensing as future work.StructFlow-HPR preserves CSI receiver-frequency topology and improves downstream HPR under limited-data conditions.Source: §V · p. 5
Loading 2609.29912v1…