SImpHAR: Impedance-Based HAR Framework

• SImpHAR is a unified physics-informed framework that synthesizes realistic synthetic bio-impedance signals from 3D human meshes for improved human activity recognition.
• It integrates a simulation pipeline with geodesic path estimation, soft-body deformation, and neural grounding to accurately map motion to impedance signals.
• The framework employs a decoupled two-stage contrastive and fine-tuning strategy leveraging both synthetic and real data, achieving significant HAR performance gains.

SImpHAR is a unified framework designed to advance impedance-based human activity recognition (HAR) by addressing the critical challenge of insufficient labeled bio-impedance data. The framework introduces a physics-informed simulation pipeline capable of generating realistic impedance signals from 3D human meshes, coupled with large-scale text-to-motion models for extensible activity synthesis. This approach is paired with a two-stage, decoupled training strategy that enables effective utilization of both synthetic and real-world data, resulting in significant improvements in HAR metrics across diverse datasets (Ray et al., 8 Jul 2025).

1. Physics-Informed Simulation Pipeline

SImpHAR constructs user-specific synthetic bio-impedance signals from 3D human meshes through three primary components: geodesic path estimation, soft-body simulation, and neural grounding, and further extends these capabilities via text-to-motion data augmentation.

1.1. Geodesic Path Estimation

SMPL-H meshes are reconstructed using methods such as OSX and interpreted as graphs $G=(V, E)$, where vertices represent mesh surface points and edges are weighted by Euclidean distances. Dijkstra’s algorithm computes the minimum-length path between electrode sites, yielding the geodesic length at each time frame $d_t$: $$d_t = \min_{p \in \mathcal{P}(s,t)} \sum_{i=1}^{|p|-1} \|v_i - v_{i+1}\|_2$$ This geodesic serves as a first-order proxy for impedance between wrist electrode pairs.

1.2. Soft-Body Deformation

To address artifacts induced by mesh discretization, a one-dimensional soft-body “beam” is simulated along the geodesic using Blender’s finite element method (FEM)-style engine. The beam’s displacement $u(\tau)$ is governed by: $$M \frac{\partial^2 u}{\partial t^2} + C \frac{\partial u}{\partial t} + K (u - u_0) = f_{\text{ext}}$$ where $M$ is the mass matrix, $K$ denotes Young’s modulus (≈10 kPa), $C$ the damping coefficient (≈0.5 N·s/m), and $u_0$ the initial geodesic. This smoothing better reflects tissue compliance and reduces high-frequency path noise.

1.3. Neural Grounding (Pose2Imp)

Personalized signal generation is achieved via a trainable dual-encoder to CNN-decoder. The system receives a window of soft-body–refined geodesic path lengths and SMPL pose parameters. The network architecture consists of two Bi-LSTM encoders (producing 512-dimensional features for path and pose), which are in turn fused by a 1D-CNN and fully connected layers to yield impedance magnitude and phase predictions. The objective is to minimize mean squared error between predicted and ground-truth impedance vectors.

1.4. Text-to-Motion Synthesis

SImpHAR leverages the PriorMDM diffusion model, conditioned on free-form text prompts, to generate plausible 3D motion sequences which are then mapped through the SMPL pipeline to synthetic impedance (“SImp”). The denoising process is guided by cross-attention on instructor-scale LLM embeddings. This mechanism enables scalable generation of proxy motion classes that may not have explicit real-world motion data, broadening the effective class space.

2. Decoupled Two-Stage Contrastive–Fine-Tuning Strategy

The SImpHARNet model is trained via a two-stage pipeline designed to address domain mismatch between synthetic and limited real data.

2.1. Stage I – Contrastive Pretraining

Synthetic impedance windows and their associated prompt embeddings are encoded by a DeepConvLSTM and a fully connected text encoder, respectively, both projecting to a 256-dimensional latent space. InfoNCE loss is used to maximize similarity between matching synthetic-text pairs: $$\mathcal{L}_{\mathrm{InfoNCE}} = -\frac{1}{N}\sum_{i=1}^N \log \frac{\exp(s(h_i^\mathrm{txt},h_i^\mathrm{imp})/\tau)}{\sum_{j=1}^N \exp(s(h_i^\mathrm{txt},h_j^\mathrm{imp})/\tau)}$$ where $d_t$0 denotes cosine similarity and τ is a temperature parameter.

2.2. Stage II – Downstream Fine-Tuning

HAR classification is performed by attaching a lightweight 1D-CNN and fully connected head to the frozen impedance encoder. An initial phase freezes the encoder, optimizing only the classifier head with cross-entropy loss; subsequently, the full network is unfrozen and jointly trained at a reduced learning rate. This “late-unfreeze” schedule preserves the semantic structure established by synthetic pretraining while adapting to individual-specific characteristics in real data.

2.3. Semantic Bridging with Proxy Motions

By generating multiple proxy motion samples per class (≈30 per 3 crafted prompts), the framework ensures embedding space alignment for both exact and analogous activities (e.g., “removing a helmet” synthesized from data for “wearing VR headset”), obviating the need for perfect label alignment in simulation and real data.

3. Empirical Evaluation and Datasets

SImpHAR is evaluated using both the newly introduced ImpAct dataset and public HAR benchmarks.

Dataset	Subjects	Classes / Tasks	Sensors & Sampling
ImpAct	10 (4F/6M)	9 (6 fitness, 3 daily)	4-electrode bioimpedance, IMU; 20 Hz
iMove	10	6 fitness	Wrist impedance+IMU; 20 Hz
iEat	10	20 eat-related	Wrist impedance+IMU; 20 Hz

Annotation in ImpAct is performed via video-based manual labeling synchronized with recordings. Experimental setup includes leave-one-user-out cross-validation, 90/10 train/validation splits per fold, early stopping, and a maximum of 300 epochs using AdamW optimization. All implementations leverage PyTorch for model training and Blender’s API for simulation.

4. Performance Analysis

SImpHAR and its ablations are assessed for both signal realism and HAR classification performance.

4.1. Neural Grounding Ablation

Input	R² (mean ± std)
3D Pose only	0.414 ± 0.018
Simulated Geodesic (d_t) only	0.821 ± 0.013
Pose + Geodesic	0.843 ± 0.025

Inclusion of soft-body–refined geodesics nearly doubles the predictive power over pose-only models, and the combined model achieves the highest fit ($d_t$1), demonstrating the value of multimodal physiological simulation.

4.2. HAR Classification Results

Dataset	Model	Accuracy	Macro F1	Relative F1 Gain
ImpAct	Baseline (real only)	0.621±0.011	0.595±0.007	–
	+Naïve SImp Aug (real+SImp)	0.618±0.016	0.581±0.012	–2.4%
	SImpHARNet (fine-tuned)	0.775±0.010	0.767±0.012	+28.9%
iMove	Baseline (real only)	0.641±0.012	0.636±0.008	–
	+Naïve SImp Aug (real+SImp)	0.743±0.014	0.739±0.010	+16.2%
	SImpHARNet (fine-tuned)	0.763±0.016	0.752±0.015	+18.2%
iEat	Baseline (real only)	0.731±0.012	0.718±0.020	–
	+Naïve SImp Aug (real+SImp)	0.669±0.015	0.651±0.013	–9.3%
	SImpHARNet (fine-tuned)	0.844±0.014	0.832±0.009	+15.8%

Improper mixing of synthetic and real data can harm performance on certain datasets, but SImpHARNet’s decoupled approach consistently provides substantial F1 gains up to +28.9%. Embedding-dimension analysis reveals that F1 saturates at 512 dimensions, and the late-unfreeze schedule achieves competitive performance even with 60% of available real data.

5. Significance, Limitations, and Prospects

The SImpHAR pipeline demonstrates several technical advantages:

• Physics-grounded geodesic and soft-body simulation considerably enhance bio-impedance signal realism.
• Large-scale text-to-motion diffusion provides scalable augmentation beyond traditional motion capture.
• Decoupled contrastive pretraining allows semantic clustering of proxy activities, mitigating the scarcity of labeled real data.
• Two-stage fine-tuning confers substantial improvements, particularly effective in low-data regimes.

Identified limitations include the necessity of per-subject neural grounding (due to unknown physiological parameters), current evaluation scope confined to wrist-to-wrist impedance, and a limited demographic pool (ages 24–35). Proposed directions include integrating physical conductivity models, broadening electrode montages and body regions, expanding demographic coverage, and generalizing the simulation core for other sensor modalities (e.g., force–pressure, capacitive, electromyography).

6. Context and Broader Implications

SImpHAR establishes a modular framework for simulating, synthesizing, and learning from bio-impedance signals, fundamentally advancing the practicality of wearable HAR in data-limited settings. Its modular, physics-informed approach, coupled with semantic augmentation strategies, indicates a generalizable methodology applicable to time-series data simulation and multimodal HAR pipelines (Ray et al., 8 Jul 2025).

A plausible implication is that similar frameworks could accelerate research across domains that require high-fidelity physiological data synthesis, surrogate modeling for rare activities, or robust cross-modal signal generation in few-shot scenarios.