Source-linked AI summary
EdgeHAR: An Edge-Native Compact Sensor Foundation Model for Human Activity Recognition
He Zhang, Siyu Yuan, Siyu Liu, Sizhen Bian, Bin Guo
TL;DR
Wearable HAR foundation models face edge-resource constraints and shifts in users, devices, sampling, and placement. EdgeHAR factorizes sensor representations and uses lightweight adaptation; experiments report transfer across heterogeneous scenarios with favorable accuracy-efficiency trade-offs.
Problem
Sensor foundation models are largely cloud-centric, while wearable HAR must accommodate edge constraints and unseen users, devices, sampling configurations, placements, and activities.
Source: §1 · p. 1Method
EdgeHAR factorizes sensor representations into activity semantics, motion dynamics, and acquisition context, then combines heterogeneous-dataset pretraining with lightweight adaptation.
Source: §1 · pp. 1–2, §2.2 · p. 2, §2.3 · p. 2Results
EdgeHAR demonstrates effective transfer across heterogeneous HAR scenarios, including unseen users, devices, placements, sampling configurations, and activity classes, with favorable accuracy-efficiency trade-offs.
Source: §1 · p. 2, §3.2.3 · p. 3, §3.2.2 · p. 3Takeaways & Limitations
The study presents structured representation learning as an effective path toward efficient and personalized ubiquitous foundation models.
Source: §4 · p. 4
Abstract
from arXiv · showhide
Sensor-based human activity recognition (HAR) is fundamental to ubiquitous and wearable computing, yet existing foundation models are largely designed for cloud-scale deployment and struggle with real-world sensing shifts, including unseen users, devices, sampling rates, and sensor placements. We present \textbf{EdgeHAR}, an edge-native compact sensor foundation model designed for wearable intelligence. Unlike conventional models that entangle activity knowledge with acquisition variations, EdgeHAR learns transferable representations by factorizing sensor signals into three latent codes: an \textbf{(i)Activity-Semantic Code} capturing reusable activity knowledge, a \textbf{(ii)Motion-Dynamics Code} modeling temporal patterns, and an \textbf{(iii)Acquisition-Context Code} representing sensor-specific variations. This disentangled design enables efficient adaptation to new users, devices, placements, and activity classes with limited target-domain data. By incorporating lightweight adaptation modules, EdgeHAR achieves foundation-model-level transferability while satisfying edge constraints in computation, memory, latency, and privacy. Experiments across heterogeneous HAR datasets demonstrate that EdgeHAR maintains competitive recognition performance under distribution shifts with substantially reduced deployment cost. EdgeHAR establishes a practical paradigm for compact, edge-first foundation models for ubiquitous sensing systems.
1 Introduction
Wearable HAR must operate within edge-device resource constraints, while existing sensor foundation models are cloud-centric and face shifts in users, devices, sampling, placements, and activities. EdgeHAR proposes disentangled representations and lightweight adaptation for transfer across heterogeneous sensing settings.
1 Introduction
- Existing sensor foundation models rely on large capacity, centralized computation, and expensive adaptation, limiting their fit for resource-constrained wearable scenarios.Pretraining and deployment also encounter sensing conditions and activity categories absent from training.Source: §1 · p. 1
- Device, sampling, and placement variations are structured sensing factors, but existing approaches often align or remove them rather than separating them from reusable activity knowledge.The paper links this entanglement to degraded representations under deployment shifts and difficult on-device adaptation.Source: §1 · p. 1
- EdgeHAR factorizes sensing representations into activity semantics, motion dynamics, and acquisition context for transferable knowledge reuse across heterogeneous environments.Its compact backbone and lightweight adaptation support edge personalization without full-model retraining or raw-data transmission.Source: §1 · pp. 1–2
- EdgeHAR demonstrates transfer across unseen users, devices, placements, sampling configurations, and activity classes with favorable accuracy-efficiency trade-offs.The paper positions this as bridging foundation-model transferability and wearable deployment constraints.Source: §1 · p. 2
2 EdgeHAR: An Edge-Native Sensor Foundation Model
EdgeHAR is a compact sensor foundation model that factorizes wearable observations and combines heterogeneous-dataset pretraining with lightweight target-environment adaptation. Its design explicitly represents acquisition variations alongside reusable activity knowledge.
2 EdgeHAR: An Edge-Native Sensor Foundation Model
- The model’s transferable representations factor sensor observations into activity semantics and acquisition-related variations for robust transfer across heterogeneous sensing conditions.The factorization distinguishes reusable activity knowledge from deployment-related variation.Source: §2 · p. 2, §2.1 · p. 2
2.1 Overview
- EdgeHAR uses multi-dataset sensor pretraining followed by edge-efficient adaptation for lightweight personalization under deployment constraints.The framework treats heterogeneous sensing conditions as structured factors rather than variations to remove.Source: §2 · p. 2, §2.2 · p. 2
2.2 Multi-Dataset Sensor Pretraining
- Pretraining combines factor-specific learning with representation consistency to produce a compact, transferable foundation representation.The consistency term preserves shared motion information, while the complementary-factor term encourages latent factors to remain complementary.Source: §2.2 · p. 2
2.3 Factorized Sensor Representation Learning
- The representation separates activity semantics, execution-dependent motion dynamics, and acquisition variations from devices, sampling configurations, and body placements.The semantic encoder seeks activity representations invariant to sensing conditions; explicitly modeling acquisition context prevents deployment-specific variation from being absorbed into semantics.Source: §2.3 · p. 2
2.4 Edge-Efficient Adaptation
- For target environments, lightweight adapters specialize the semantic representation using limited labeled data instead of full-model fine-tuning.Only a small number of parameters are updated, and deployment retains the compact recognition pathway for efficient wearable inference.Source: §2.4 · p. 2
3 Evaluation: Towards Practical Edge Foundation Intelligence
EdgeHAR is evaluated for transfer under dataset, user, and placement shifts, with 2% target-domain samples used for fine-tuning. The evaluation also examines edge deployment costs and adaptation with limited target data.
3.1 Experimental Setup
- Cross-dataset, cross-person, and cross-position evaluations use only 2% of target-domain samples for fine-tuning.The setup defines these shifts as transfer to unseen datasets, users, and sensor placements, respectively.Source: §3.1 · p. 2
3.2 Transferability under Sensing Shifts
- Across multi-source and single-source transfer, EdgeHAR outperforms baselines on unseen target datasets, with single-source gains over zero-shot transfer using 2% target data.Multi-source pretraining uses three heterogeneous datasets; single-source evaluation adapts from one dataset to unseen targets.Source: §3.2.3 · p. 3, Table 2 · p. 3, §3.2.2 · p. 3
- On unseen users and placements, EdgeHAR-FS outperforms transfer approaches, while placement evaluation also reports strong zero-shot performance and further few-shot gains.Cross-person transfer achieves the highest average accuracy across datasets; cross-position testing uses leave-one-position-out transfer on Shoaib.Source: §3.2.2 · p. 3, §3.2.4 · p. 3, Table 4 · p. 4
3.3 Edge Efficiency
- The deployment version uses 1.086M parameters and 4.14 MB storage, while requiring 18.42 MB memory, compared with the full model’s 1.41M parameters and 5.41 MB storage.Source: §3.3 · p. 4
3.4 Adaptation under Limited Target Data
- With 100 samples per emerging activity, EdgeHAR reaches 91.86% Macro-F1; performance also rises as target fine-tuning ratios increase from 0% to 15% across datasets.The emerging-activity evaluation uses PAMAP2 and tests adaptation beyond the pretraining label space.Source: §3.4.1 · p. 4, Table 7 · p. 4, §3.4.2 · p. 4
4 Discussion and Conclusion
EdgeHAR combines disentangled representations and lightweight adaptation for robust, personalized wearable HAR across sensing shifts, with a small footprint for real-time on-device inference.
4 Discussion and Conclusion
- Experiments across unseen users, devices, placements, and sampling rates confirm EdgeHAR’s robustness.Its small footprint supports real-time on-device inference.Source: §4 · p. 4
- EdgeHAR disentangles activity semantics from acquisition variations and uses lightweight adaptation to enable transferable edge intelligence.Source: §4 · p. 4
- The study presents structured representation learning as an effective path toward efficient and personalized ubiquitous foundation models.Source: §4 · p. 4