This work introduces RAG-HAR+, a retrieval-first and cost-optimized extension that strengthens retrieval while reducing dependence on LLM-based inference, and extends the RAG-HAR mobile prototype to demonstrate the practical feasibility of retrieval-first, LLM-assisted HAR in mobile sensing scenarios.
Abstract
Human Activity Recognition (HAR) from wearable sensors supports applications in healthcare, rehabilitation, fitness tracking, and smart environments. Yet, existing deep learning approaches require dataset-specific training, large labeled corpora, and repeated adaptation to new sensor settings or activity taxonomies. Retrieval-Augmented Generation for Human Activity Recognition (RAG-HAR) addresses this by framing HAR as a training-free, retrieval-augmented task, in which statistical descriptions of sensor windows are used to retrieve similar labeled examples that guide LLM-based classification. We introduce RAG-HAR+, a retrieval-first and cost-optimized extension that strengthens retrieval while reducing dependence on LLM-based inference. RAG-HAR+ uses an offline Retrieval Designer Agent to design dataset-specific feature groups from a diverse pool of motion descriptors, enabling sensor windows to be compared using features better aligned with dataset-specific activity patterns. During inference, RAG-HAR+ uses majority voting over retrieved neighbors for samples with strong retrieval evidence and defers only uncertain cases to an LLM-based Ambiguity Resolver Agent. Across six HAR benchmarks, RAG-HAR+ maintains competitive or improved performance while reducing LLM usage, token consumption, and inference time. We further extend the RAG-HAR mobile prototype to demonstrate the practical feasibility of retrieval-first, LLM-assisted HAR in mobile sensing scenarios.
A novel lightweight cross-domain few-shot sensor-based HAR network (CFSH-Net) is proposed for cross-domain activity recognition with limited labeled samples, which demonstrates strong cross-user generalization on PAMAP2 and USC-HAD, and stable cross-dataset transfer when trained on OPPORTUNITY and evaluated on four other datasets.
Hao Zheng, Hongji Xu, Fei Gao et al.· IEEE journal of biomedical a...· 0 citations
Wi-Fi-based human activity recognition (HAR) using channel state information (CSI) provides a nonintrusive and device-free sensing solution for smart cities, healthcare monitoring, and smart homes. However, recognition performance often degrades when models trained on limited users are applied to unseen users due to variations in body shape, posture, movement style, and surrounding conditions. To address this cross-user robustness challenge, this article proposes CUTA-HAR, a Cross-User Temporal Attention Network for Wi-Fi CSI-based HAR. CUTA-HAR combines multiuser supervised training with an attention-based bidirectional LSTM (BiLSTM) to capture informative temporal CSI patterns from multiple training users, without requiring data from the unseen test user during training. Experimental evaluations on a self-collected multiuser CSI dataset show that CUTA-HAR consistently outperforms representative sequence modeling baselines under a leave-one-user-out evaluation protocol, achieving average test accuracy improvements of 2.0%–6.7%. Action-level analysis further shows that structured activities can be recognized reliably, while complex activities such as fall and pickup remain challenging due to larger cross-user motion variations. These results indicate the effectiveness of attention-guided temporal modeling for improving cross-user robustness in Wi-Fi CSI-based HAR.
Fucheng Miao, Osamu Takyu, Shan Lin et al.· IEEE Internet of Things Jour...· 0 citations
Fine-grained wrist activity recognition can support applications such as procedural step guidance and context-aware assistance, yet acquiring labeled data for every new task, user, and activity granularity remains a bottleneck. We present TransfHAR, a self-supervised wrist IMU framework for on-demand, fine-grained activity recognition by learning transferable motion priors from global, unlabeled activities. We show that self-supervised pretraining on coarse wrist IMU activities (e.g., sitting, walking, exercise) learns motion structure rich enough to transfer to fine-grained manipulative, gestural, and procedural activities (e.g., snapping, stirring, waving) that are absent from pretraining. We implement TransfHAR as a real-time smartwatch application that lets users define and expand their own activity set for personalized recognition from only a few demonstrations. Across three offline cross-dataset evaluations, TransfHAR matches or exceeds fully supervised baselines that use complete label sets with equal or additional sensor channels, by 6.2 balanced-accuracy points on average. In an in-lab study with 10 participants each performing seven novel wrist activities, TransfHAR reaches 86.7% balanced accuracy across participants with five examples per class and 90.4% when updated from a single one-minute recording per class. These results indicate that broad self-supervised wrist pretraining provides an effective foundation for on-demand fine-grained activity recognition.
Aidan Bradshaw, Riku Arakawa, Xin Liu et al.· 0 citations
Smart home environments provide a practical basis for privacy-preserving activity monitoring in elderly-care and independent living settings, where ambient sensors such as motion detectors and door contacts can observe daily routines without requiring any action or device from residents. However, ambient-sensor-based human activity recognition (HAR) faces a persistent challenge: fine-grained activity labels commonly used in benchmark datasets, such as Cook Breakfast, Cook Lunch, and Cook Dinner, often exceed the discriminative capacity of sparse environmental sensors, introducing label ambiguity and fragmenting the training data available per class. To address this issue, this paper proposes a sensor-aligned activity taxonomy that reorganizes activity classes according to sensor distinguishability rather than semantic granularity. The taxonomy is integrated with temporal feature engineering and a personalized LightGBM-based recognition pipeline optimized for edge deployment. Experiments on 25 households from the CASAS smart home dataset show that the proposed approach improves mean recognition accuracy from 74.83% to 82.45% and increases the F1-score for the clinically relevant activity Take Medicine from 0.42 to 0.61. These results suggest that aligning activity label design with sensor observability can meaningfully improve recognition accuracy without increasing model complexity.
Le Bao Ngoc Tran, Duc Dat Pham, Huu-Sy Le et al.· 2026 11th International Conf...· 0 citations
This study reviews pre-2018 developments and proposes an improved HAR framework incorporating advanced feature engineering, sensor fusion, and ensemble learning techniques, which shows strong potential in healthcare, fitness, and smart environments.
Silvia Diallo, F. Z. Idrissi· International Journal of Mod...· 0 citations