Aug 2026· Journal of King Saud University: Computer and Information Sciences· Vol 38· 0 citations· 35 references
TL;DR
MST-HDQ, a compact sequence-code learning framework that co-designs a sensor-aware large-kernel temporal encoder, wearable-sequence hierarchical attentive aggregation, a hash-discriminative quantization head, and a class-robust training objective, is proposed.
Abstract
Sensor-based human activity recognition (HAR) from wearable devices is fundamental to applications such as mobile health monitoring, smart sports training, rehabilitation assessment, and context-aware computing. Wearable sensor sequence recognition requires representations that are temporally expressive, compact, and robust to class imbalance. Existing sensing systems typically treat temporal modeling and compact representation learning as separate design problems: temporal architectures often terminate in dense classifier logits, whereas hash-based objectives operate on features that have already been formed. This decoupled design limits the discriminability of learned activity codes, particularly for minority classes in imbalanced datasets. We propose MST-HDQ, a compact sequence-code learning framework that co-designs a sensor-aware large-kernel temporal encoder, wearable-sequence hierarchical attentive aggregation, a hash-discriminative quantization head, and a class-robust training objective. The temporal encoder combines multi-scale local branches with depthwise large-kernel temporal blocks to capture both short transient motions and longer activity phases, while the aggregation module fuses attentive temporal pooling with multi-depth feature summaries to select informative temporal moments across encoder depths. The quantization head learns compact prototype-discriminative sequence codes through scalar codebook discretization and normalized prototype classification. Experiments on OPPORTUNITY, PAMAP2, WISDM, and UniMiB-SHAR under the Hash-MMDC protocol show consistent macro-F1 improvements over reproduced baselines, with paired 50-seed Wilcoxon tests showing p<0.001\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p < 0.001$$\end{document}. Controlled comparisons with additional re-implemented baselines, grouped ablations, hash-space diagnostics, sensitivity studies, and stricter segment-grouped and subject-independent splits further support the robustness of the gains. MST-HDQ preserves compact 64-dimensional codes while requiring only 0.622M–0.813M trainable parameters, making it suitable for efficient wearable sensing deployments. The results demonstrate that jointly optimizing temporal evidence extraction and code-space geometry yields more class-complete hash structures than treating these objectives independently.
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
Human activity recognition (HAR) plays a critical role in intelligent wireless sensing and mobile edge computing. Compared with traditional vision-based and wearable-based approaches, WiFi channel state information (CSI) enables privacy-preserving and device-free activity perception. CSI encodes human-induced channel variations that serve as a natural basis for activity-oriented semantic analytics over wireless networks. However, existing WiFi CSI-based methods suffer from weak cross-scene generalization, high model complexity, and poor multi-task collaboration. To address these issues, this paper proposes UniSense-CSI, a unified multi-task framework that jointly learns dynamic gesture recognition, static posture classification, and fall detection. Specifically, the proposed framework converts CSI signals into pseudo-RGB images, extracts spatio-temporal features using a customized ConvNeXt backbone, and leverages an improved PerceiverIO module to compress high-dimensional features into a compact latent space. Based on the shared representation, task queries and adapters are introduced to enable parallel multi-task inference within a common architecture. Experiments on public datasets demonstrate that the proposed framework achieves accuracies of 99.64%, 99.66% and 96.09% on the three tasks, respectively, while maintaining low inference latency and favorable edge-deployment capability.
Fall detection for wearable health monitoring must combine subject-independent accuracy, low false-alarm risk, real-time response, and multi-day battery operation. Deep learning can capture fall dynamics, but recurrent or long-window models often increase memory access, inference latency, and energy use on microcontroller-class devices. This paper introduces CNN-FIFO-LSTM, a compact TinyML architecture that decouples spatial and temporal processing for real-time fall detection. A convolutional neural network maps each four-channel inertial window to a 32-dimensional embedding, a first-in-first-out (FIFO) buffer stores recent embeddings, and a lightweight LSTM classifies the resulting temporal sequence. Applying recurrence to compressed embeddings rather than raw sensor windows preserves temporal context while reducing both the sequence-model input volume and computational resource use. On KFall, leave-one-subject-out (LOSO) validation achieves 98.85% accuracy, 97.04% precision, 98.95% recall, and a 97.98% F1-score. When trained on KFall and evaluated on SisFall, the model maintains 97.4% accuracy and a 96.7% F1-score. The selected floating-point ESP32-S3 deployment requires 227 KB for the models, 573 KB of firmware flash, and 132.3 KB of RAM, with 48.9 ms of model inference, 56.6 ms of the full cycle, and 5.3 days of estimated continuous operation from a 1000 mAh battery. A mixed-precision post-training quantization variant, using an int8 CNN and a float32 LSTM, reduces inference latency to 28.3 ms and extends estimated battery life to 7.8 days, with a modest reduction in F1-score to 97.67%. Together, these results show that CNN-FIFO-LSTM achieves a practical trade-off among detection performance, response time, and battery life in evaluations using public datasets and prototype hardware.
Duan Luong Cong, Cuong Chu Van, Anh Pham Hoang et al.· IEEE Access· 0 citations
The proposed Joint Embedding Predictive Architecture framework designed to learn robust and generalizable representations from unlabeled datasets demonstrates superior generalization on minority, high variance transitional activities such as sit-to-stand and sit-to-lie where supervised learning tend to overfit due to limited support.
Mohd Halim Mohd Noor, AbdulRahman M. A. Baraka· arXiv.org· 0 citations
Human Activity Recognition (HAR) systems using deep learning have shown significant promise; however, deploying such models on edge computing systems remains challenging due to constraints in inference latency, memory footprints and computational capacity. This study proposes a lightweight, end-to-end patch-based Transformer architecture designed for efficient HAR for resource constrained edge environments, evaluated through an architectural ablation study and multiple quantization strategies using TensorFlow Lite. We conducted extensive experiments by varying patch lengths and Transformer encoder depths in order to evaluate the impact on classification accuracy, latency and model complexity. The results demonstrate that moderate temporal patching achieves an effective balance between temporal representation learning and computational efficiency. Experiments have been conducted on two widely open-source benchmark datasets, WISDM and PAMAP2, where raw sensor signals are pre-processed through segmentation, normalization, and overlapping windowing before being fed into a compact Transformer model with patch embedding and multi-head self-attention for feature extraction. The trained models are converted into FP32, FP16, INT8 dynamic range, and INT8 full formats to evaluate trade-offs between different metrics such as accuracy, model size, and inference latency. The baseline model achieves 96.77% accuracy on WISDM and 98.48% on PAMAP2, with consistently high macro F1-scores. Among all quantization variants, TFLite-INT8-Dynamic reduced model size by 95.55% and 89.80% for WISDM and PAMAP2, respectively and drastically improved inference latency by around 99% for both datasets with minimal accuracy degradation. These findings demonstrate that the proposed patch-based Transformer model for HAR has achieved an effective balance between recognition performance and computational efficiency, which indicates its strong deployment feasibility for resource constrained edge environments through TensorFlow Lite benchmarking and can provide a scalable solution for edge intelligence applications.
Wearable inertial measurement units (IMUs) offer an objective, low-cost basis for Parkinson’s disease (PD) assessment, but multi-task clinical protocols yield heterogeneous recordings across body locations and small cohorts, and it is unclear whether such data can support reliable PD detection without training deep models from scratch. We therefore ask whether a motion-pretrained representation transfers to this setting, and quantify how much of the discriminative signal it supplies. Each subject is represented by five task-level motion embeddings, one per clinical task, produced by a frozen pretrained spatio-temporal graph convolutional network (ST-GCN) that fuses the thirteen body-worn sensors into a whole-body embedding; a three-layer Transformer with validity-mask weighting aggregates these tokens for binary PD-versus-control classification on the WearGait-PD cohort (181 subjects: 100 PD, 81 controls). Under a leakage-free nested protocol with repeated subject-disjoint stratified 5-fold cross-validation (5 seeds; 25 estimates per model) and paired significance testing, the model attains a balanced accuracy of 0.834 ± 0.087, macro-F1 of 0.842 ± 0.094, and AUC of 0.842 ± 0.103. It leads six classical baselines and a spectrogram-CNN on accuracy-based metrics, though random forest, gradient boosting, and the spectrogram-CNN edge ahead on AUC; after correction for fold correlation, none of these between-model differences is significant. The one robust finding is a transfer effect: replacing the pretrained encoder with a random one of identical architecture lowers balanced accuracy by 15.5 points when frozen (p = 0.043) and 20.4 when trained end-to-end (p = 0.014). Discrimination is preserved under 1:1 age matching (0.846) and across both genders, so it is not explained by age imbalance. Motion-pretrained skeletal encoders thus supply the majority of the discriminative signal, while the aggregator contributes gains inseparable from noise at this cohort size.
H. Herath, Nuwan Madusanka Vithanage, Chaminda Hewage et al.· Bioengineering· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
