Search Results (463)

Sensor-based human activity recognition (HAR) plays a fundamental role in healthcare monitoring, sports analytics, and ambient-assisted living. Although deep learning has substantially advanced HAR performance, two practical issues still limit its real-world deployment: (i) the distribution shift caused by changes in users or sensor placements can degrade generalization, and (ii) the quadratic $\mathcal{O}\left(L^2\right)$ complexity of standard self-attention hinders efficient long-sequence modeling on resource-constrained wearable devices. To address these issues, we propose DSHformer, which is an accuracy-oriented HAR framework that combines compact channel–temporal encoding with locality-sensitive hashing (LSH)-based attention. Specifically, DSHformer (i) employs a low-parameter patch-based graph-attention encoder to jointly model latent relationships among sensor channel–temporal dynamics; (ii) introduces a trainable prototype pool together with a multi-layer decomposition network to improve intra-class compactness and inter-class separability via prototype alignment; and (iii) introduces a decomposition-stable LSH-based attention mechanism tailored for HAR, whose core design couples prototype-guided feature decomposition with locality-sensitive hashing to ensure that semantically related tokens remain consistently grouped in the same hash bucket even after decomposition-induced attenuation. The mechanism thereby operates at $\mathcal{O}(LlogL)$ attention complexity on longer sensor sequences. Extensive experiments on five public benchmarks (WISDM, UCI-HAR, PAMAP2, Opportunity, and UniMiB-SHAR) show that DSHformer achieves accuracies of 98.6%, 93.7%, 98.4%, 88.5%, and 96.6%, respectively, achieving competitive or superior performance compared with both Transformer variants and HAR-specific baselines under the adopted benchmark protocols. Ablation studies further confirm the complementary contribution of each component. Full article

Wearable-sensor human activity recognition (HAR) plays a key role in health monitoring, elderly care, and human–computer interaction. Deep learning dominates the field, but two limitations remain. CNNs with fixed kernels cannot capture cross-scale temporal events such as gait cycles and postural transitions in a single layer, and softmax attention on small sensor datasets is often diluted by common-mode background responses across the sequence. We propose DMSCNet, an end-to-end framework with two modules. The Dilated Multi-Scale Branch Block (DMSB) combines a shared bottleneck, parallel dilated convolutions, a pooling bypass, and SE-based channel recalibration to widen the temporal receptive field under a controlled parameter budget. The Contrastive Temporal Attention (CTA) module adopts a dual-path differential design, in which the two paths learn overlapping but non-identical attention patterns and their subtraction suppresses shared low-level responses while preserving the discriminative positions each path locks onto, encoded with opposite signs. DMSB and CTA are cascaded into a DMSC Block and stacked residually. On UCI-HAR, USC-HAD, and RealWorld, DMSCNet reaches F1-scores of 97.65%, 91.80%, and 99.05%, outperforming nine baselines. Ablations confirm that SE acts along the channel axis and CTA along the temporal axis, and visualization reveals a dynamic–static dichotomy together with a signed bipolar encoding pattern produced by the dual-path subtraction. Full article

Human activity recognition—HAR—plays a crucial role in the lives of patients battling neurodegenerative diseases. These debilitating conditions, such as Alzheimer’s or Parkinson’s, affect individuals’ ability to perform daily tasks autonomously and safely. HAR technology offers an invaluable solution by enabling real-time monitoring and assistance, helping to maintain independence and quality of life for patients. Additionally, this technology provides a valuable data source for doctors and caregivers, allowing for more precise and personalized care, which can make a difference in managing and treating these neurodegenerative diseases. The objective of this review is to identify the contribution of Transfer Learning and Reinforcement Learning in supporting the processes of daily activity recognition, thus enhancing the quality of life for patients. As this is a trending topic, the literature surrounding it is quite dispersed, which is why this review aims to present the current line of research in this field. To carry out this analysis, the science tree paradigm was used, which establishes two fundamental stages of analysis. The first is delimited by scientometrics, where the leading countries in the application of such technologies can be identified. This review highlights the evolution in the use of transfer learning and reinforcement learning in HAR in the healthcare field, where these techniques have significantly improved the accuracy and adaptability of real-time monitoring systems. The studies reviewed indicate that transfer learning has allowed models to adapt to data variations without requiring large volumes of manual labeling, which is essential in clinical and patient monitoring contexts. Additionally, reinforcement learning has optimized decision-making in complex scenarios, enabling activity recognition systems to dynamically adjust monitoring parameters, enhancing detection and response to critical or unusual activities in multi-user environments. These advances demonstrate that, by integrating these approaches, greater personalization and robustness can be achieved in human activity recognition, thereby improving the quality of life for patients in clinical settings. Full article

Human activity recognition (HAR) using Wi-Fi channel state information (CSI) offers a privacy-preserving and contactless sensing modality suitable for smart homes, healthcare monitoring, and pervasive mobile Internet of Things (IoT) environments. However, existing CSI-based HAR approaches often suffer from computational inefficiency, high latency, and limited feasibility on resource-constrained embedded platforms. This work presents STAR (Sensing Technology for Activity Recognition), an edge AI-optimized framework that integrates lightweight temporal modeling, adaptive signal processing, and hardware-aware co-optimization to enable real-time, energy-efficient HAR on low-power embedded devices. STAR employs a streamlined three-layer Gated Recurrent Unit (GRU) architecture that reduces model parameters by 33% compared to conventional Long Short-Term Memory (LSTM) designs while maintaining strong temporal modeling capability. To enhance signal quality, STAR incorporates a multi-stage pre-processing pipeline consisting of median filtering, an eighth-order Butterworth low-pass filtering, and empirical mode decomposition (EMD) to denoise CSI amplitude measurements and extract stable spatial-temporal features. For on-device deployment, the system is implemented on a Rockchip RV1126 processor equipped with an embedded Neural Processing Unit (NPU) and interfaced with an ESP32-S3 CSI acquisition module. Experimental results demonstrate a mean recognition accuracy of 93.52% across seven activity classes and 99.11% for human-presence detection using a compact 97.6k-parameter model. INT8-quantized inference achieves a processing throughput of 33 MHz with only 8% CPU utilization, achieving a six-fold improvement in inference speed over CPU-based execution. With sub-second response latency and low power consumption, the system ensures real-time, privacy-preserving HAR, offering a practical, scalable solution for mobile and pervasive computing environments. Full article

Wireless Sensor Networks (WSNs) are increasingly evolving toward intelligent distributed systems in which local sensing, on-device inference, and collaborative model training are becoming central to scalable Internet of Things (IoT) deployments. However, the practical adoption of Federated Learning (FL) in WSN-oriented environments remains constrained by three major challenges: limited and unevenly depleted node energy, heterogeneous non-IID local data distributions, and variable client reliability during collaborative training. This paper proposes a Trust- and Energy-Aware Federated Learning (TEA-FL) framework specifically designed for resource-constrained WSN settings, in which client participation and server-side aggregation are jointly guided by residual energy estimates and dynamically updated trust scores. The proposed method prioritizes reliable, energy-efficient sensor nodes while reducing the impact of weakly aligned or low-quality local updates during global aggregation. The framework is evaluated on two representative WSN/IoT-oriented proxy benchmarks, Human Activity Recognition (HAR) and UNSW-NB15 intrusion detection, under both IID and Dirichlet-based non-IID federated partitions. Under non-IID HAR partitioning, TEA-FL improved final accuracy from 0.6752 with FedAvg to 0.7636 and final Macro-F1 from 0.5623 to 0.7185. On the more challenging non-IID UNSW-NB15 benchmark, TEA-FL achieved the highest final Macro-F1, 0.3711, compared with 0.3230 for FedAvg and 0.3323 for the trust-only baseline, although with a lower final accuracy. These results indicate that TEA-FL is particularly useful when final-round robustness, class-balanced behavior, and client sustainability are more relevant than maximizing a single peak intermediate accuracy value. Additional ablation and unreliable-client experiments further show that the trust–energy-aware aggregation component is particularly influential and that TEA-FL can improve behavior under selected low-quality participation scenarios, although it should not be interpreted as a complete Byzantine-robust defense. Overall, the findings suggest that jointly modeling update consistency and residual energy offers a practical, lightweight pathway toward more dependable and sustainable federated intelligence in next-generation WSN and IoT deployments. Full article

Human activity recognition (HAR) from wearable sensors increasingly faces a dual bottleneck: obtaining labels is expensive, and the labeled subset is often class-imbalanced and redundant. We address this problem with a budget-aware class-balanced active learning framework, termed CCUR-M, that closes the loop between adaptive class balancing, hybrid batch querying, and lightweight retraining. At each round, the labeled subset is rebalanced toward a median target class size through cluster-preserving majority undersampling and minority-class conditional synthesis, after which a hybrid query score combines minimum-confidence uncertainty with cluster-centered representativeness under a round-dependent budget weight. An XGBoost classifier is retrained on the rebalanced set, and the procedure is iterated until the annotation budget is exhausted. We evaluate the method on three public wearable HAR benchmarks with different difficulty profiles: PAMAP2, OPPORTUNITY, and USC-HAD. CCUR-M achieves the best final Macro-F1 on all three datasets, reaching 0.9574, 0.6780, and 0.6128, respectively. The largest final and average gains over the strongest baseline occur on OPPORTUNITY (+0.1205 final, +0.0629 average), while USC-HAD reveals a later-stage rather than early-stage advantage. Ablation experiments show that no single module explains the overall gain; instead, balancing, uncertainty, and representativeness act synergistically, with the full loop outperforming the base variant by +0.1243, +0.1638, and +0.2143 on PAMAP2, OPPORTUNITY, and USC-HAD. These results support a mathematically interpretable view of active learning for imbalanced wearable time series: the key benefit arises from coupling distribution correction and query design within the same budgeted training loop. Full article

Driven by the profound convergence of the Internet of Things (IoT) and ubiquitous computing, wearable multifunctional sensors have emerged as a key technology for high-precision human activity recognition (HAR). Advancements in novel materials and flexible electronics have propelled the evolution of these sensors, enabling advances in decoupling heterogeneous signals, enhancing system robustness, and expanding environmental perception. This review systematically examines the frontier research on wearable multifunctional sensors for HAR. We provide an in-depth analysis of three core architectural design paradigms: architecture-level integration, which relies on physical spatial isolation for hardware-level signal decoupling; monolithic integration, which strives for extreme spatial compactness and spatiotemporal signal consistency; and the emerging intrinsically multifunctional design, which leverages novel stimuli-responsive materials for the intrinsic orthogonal discrimination of multidimensional signals. Furthermore, we delineate the diverse application scenarios of these highly integrated sensing platforms across medical rehabilitation, sports science, human–computer interaction (HCI), and daily behavior perception. Finally, this article discusses the critical challenges currently confronting this technology and outlines its future development prospects. Full article

To improve operational safety in confined spaces, this study proposes an intelligent safety monitoring framework that utilizes multimodal data from wearable devices. The framework comprises two core components: a human activity recognition (HAR) module and a bio-inspired adaptive multimodal decision fusion (BA-MDF) module. The HAR module processes accelerometer and gyroscope data through an enhanced FFT–LSTM architecture that integrates time- and frequency-domain features for real-time activity classification. The BA-MDF module, inspired by biological multisensory integration mechanisms—particularly the inverse effectiveness principle observed in the superior colliculus—evaluates contextual risk by adaptively fusing HAR outputs, heart rate variability, and geospatial constraints without additional computational overhead. Experimental testing demonstrated 92.4% overall HAR accuracy and 94.3% identification accuracy for emergency scenarios under a simulated sensor degradation environment. These results validate the framework’s effectiveness in mitigating risks from anomalous events in visually constrained environments. Full article

Human activity recognition (HAR) plays a multi-dimensional supporting role in the medical field, providing strong technical support for various aspects such as disease prevention, diagnosis, treatment and rehabilitation. However, the use of traditional federated learning to deploy HAR models on edge devices is not ideal because of the heterogeneity of hardware and data. To solve this problem, this paper introduces a personalized HAR method, which can remove the outlier nodes and cluster hierarchically. In this study, the cosine similarity of local model parameters is calculated, and the clustering of dynamic clients is realized. In the study, the normalized training loss evaluation mechanism is introduced to identify and eliminate outlier nodes, and the robustness of the system is enhanced. In the study, the collaborative training method is adopted to meet the personalized needs of users and improve the generality of the model. The proposed method achieves an average recognition accuracy of 92.94% and an F1 score of 91.28% on four public datasets, demonstrating that the method put forward in this paper can reduce the negative impact of data heterogeneity, improve the efficiency of convergence, and produce good recognition performance for the development of the Internet of Things. Full article

Using attribution-based priors to improve the temporal interpretability and robustness of dynamic neural networks provides a computationally efficient method that does not alter the model structure during inference. We explore explanation-guided training for timeseries classification through the introduction of attribution-sensitive loss terms that serve as regularizers for the evolution of input relevance over time. The main contributions are the Temporal Relevance Smoothness Index (TRSI) and a ratio-based loss that reduces irregular step-to-step changes in channel-aggregated absolute relevance. TRSI is compared against temporal total-variation penalties computed using Layer-wise Relevance Propagation Total Variation (LRP-TV) and Integrated Gradients Total Variation (IG-TV). Experiments on a controlled three-class subset of the Korean University Human Activity Recognition (KU-HAR) dataset using a finite impulse response neural network (FIRNN) show that TRSI yields the strongest smoothness improvement, reducing the total variation of the aggregated relevance signal from 0.768 to 0.447 (41.8%), compared with 0.667 (LRP-TV) and 0.677 (IG-TV). Robustness tests indicate a clear advantage for TRSI under impulsive and white Gaussian test-time noise. Full article

Wearable inertial sensors produce multi-axis motion signals with rich spatial and temporal structure. Existing deep-learning pipelines for human activity recognition (HAR) rarely tackle three issues jointly: explicit modeling of the body-part grouping of multi-location inertial channels, bidirectional temporal modeling at linear-time cost, and dynamic, time-varying attention for non-stationary motion. We aim to close these three gaps within a single architecture. To this end we propose Dynamic-Attentive Selective Mamba (DASM), which combines three components: Group-Aware Convolutions (GroupConv) for body-part-aware local features, a Bidirectional Mamba (BiMamba) module for linear-time forward and backward temporal context, and a Dynamic CBAM (DCBAM) that produces per-timestep channel and spatial attention for non-stationary windows. On the UCI Daily and Sports Activities dataset (19 classes, 8 subjects), under stratified segment-level 5-fold cross-validation (3 seeds, 15 runs/model), DASM reaches 99.89% accuracy and F1, a 0.11% gain over CNN-BiGRU-CBAM and 0.50% over Multi-STMT; under leave-one-subject-out (LOSO), it reaches 89.34%, 1.69% above the strongest baseline. The 10.55% drop under LOSO shows that segment-level results overestimate cross-subject generalization. Ablations show small but statistically detectable gains (Cohen’s $d\in [0.4,0.7]$ per module, $d\ge 1.5$ full-vs-baseline). We therefore position the contribution as a structured architecture within a near-saturated benchmark; broader deployment claims require multi-dataset subject-independent validation. Full article

This paper presents a robust and efficient mmWave radar-based human activity recognition (HAR) framework optimized for practical real-time indoor deployment. Addressing computational inefficiencies and limited recognition scopes in existing systems, the framework introduces two core contributions: Multi-class Spatio-Temporal Network (MuST-Net), a lightweight, multi-class network, and an online detection process for enhanced temporal stability. MuST-Net utilizes a hybrid 2D convolutional neural network and temporal convolutional network architecture to recognize seven distinct classes, significantly broadening the system’s recognition repertoire. The online detection process implements a novel sliding-window post-processing chain that employs an activity-buffering mechanism, which maintains temporal continuity and effectively suppresses spurious detections at activity boundaries. Experimental results demonstrate the superior performance of our unified framework, attaining over 98.6% accuracy for multi-class classification by MuST-Net and achieving at least 97% accuracy for activity detection and a crucial 100% recall for fall detection. Robustness is validated across three distinct indoor environments and nine subjects—with two of the three sites entirely unseen during training—confirming strong generalization under installation, environment, and subject variations. Full article

Complex human activities consist of sequential, simple limb movements, acting as impulse responses from the motor system. In short-term human activity recognition (ST-HAR), the inherently brief observation window results in non-stationary signals and “information starvation,” breaking the time-translational symmetry of kinetic signals. Moreover, traditional Variational Mode Decomposition (VMD) and Hilbert Transform (HT) suffer from suboptimal decomposition levels (K) and spectral asymmetry. This paper proposes an improved VMD-HT framework to enhance feature extraction from short-term Inertial Measurement Unit (IMU) signals. First, an instantaneous-frequency-driven adaptive VMD method is developed to mitigate mode mixing by automatically determining the optimal K. Second, an information-enhanced instantaneous energy density (IEIE) feature is introduced. By fusing kinetic energy from both positive and negative frequency domains, this feature restores the spectral symmetry of the energy representation, precisely quantifying fine motion variations and compensating for information loss caused by the limited temporal span. Experimental results on PAMAP2, WARD, and a self-collected dataset, NOITOM, demonstrate the method’s effectiveness. With a 0.5 s window, the proposed model achieves outstanding recognition accuracies of 93.60%, 96.41%, and 97.22%, respectively, outperforming state-of-the-art approaches in capturing transient short-term information. Full article

Human Activity Recognition (HAR) from inertial sensors must run continuously on battery-powered wearables under tight latency, memory, and energy budgets. While tiny Transformers can be effective on inertial time series, end-to-end co-design across quantized inference and heterogeneous low-power platforms remains underexplored. We present STORM (Small Transformer for On-node Recognition of Motion), a deployment-oriented 19.7k-parameter 1D Transformer co-designed with X-HEEP, an open-source low-power single-core RISC-V SoC, and a tightly coupled streaming CGRA for nonlinear primitives (e.g., softmax). We build a cross-source 8-class benchmark by harmonizing 3 public datasets under a stringent, deployment-aligned protocol that exposes both cross-subject and cross-source shift. Using 1.280 s windows with 0.640 s stride, the protocol models continuous on-node HAR under cross-dataset generalization. After quantization-aware training and INT8 C inference export, STORM achieves 0.799/0.801 accuracy/macro-F1 on this benchmark. Deployed on an FPGA prototype of X-HEEP with the streaming CGRA backend, STORM requires 67.4 ms per inference at 100 MHz, while activity-based power analysis estimates a total inference energy of 632.4 $\mathsf{\mu }$J, satisfying the stride-driven real-time constraint. These results support the practical viability of compact attention-based HAR on low-power wearable-class embedded platforms. Full article

Wearable sensor-based human activity recognition (HAR) has gained growing significance in healthcare monitoring and assisted living systems. Although considerable advances have been made in classifying dynamic movements, stationary activities—such as sleeping, driving, and watching TV—remain difficult to distinguish owing to their weak sensor signatures and limited discriminative cues. This paper presents the multi-scale hybrid attention temporal network (MHAT-Net), a deep learning framework whose key architectural novelty lies in the parallel (non-sequential) dual-pathway temporal modeling: a BiGRU branch and a transformer encoder branch operate simultaneously on the same spatially encoded representation, combined via a learnable attention-based fusion module. This design targets the underexplored problem of distinguishing stationary activities from weak inertial sensor signatures. The architecture is built upon three integrated components: (1) a multi-branch CNN with kernel sizes three, five, and seven combined with channel attention for adaptive spatial feature extraction across multiple temporal scales; (2) parallel bidirectional gated recurrent unit (BiGRU) and transformer encoder pathways for jointly capturing short-range sequential patterns and long-range temporal correlations; and (3) an attention-driven fusion module that adaptively weights the outputs of both temporal branches. The model was assessed on a publicly available benchmark comprising three motionless activity categories collected from 25 participants via smartphone sensors. In 5-fold cross-validation, MHAT-Net attained 97.42% (±4.69%) accuracy with accelerometer data and 92.31% (±0.31%) with gyroscope data, substantially exceeding the accuracies of five baseline architectures: CNN, LSTM, BiLSTM, GRU, and BiGRU. Ablation experiments identified multi-scale spatial feature extraction as the most influential module (2.21–2.47% contribution), followed by the hybrid temporal modeling components. Cross-modality analysis confirmed that accelerometer signals yielded richer discriminative content for stationary activities, while MHAT-Net sustained consistent performance across both sensor types. The proposed integration of multi-scale spatial encoding, hybrid temporal modeling, and multi-level attention gives MHAT-Net the ability to reliably detect subtle activity-specific patterns, establishing a new benchmark in wearable sensor-based recognition for comprehensive daily behavior monitoring. Full article