Stacking ensemble-based predictive framework for human activity and postural transition recognition using wearable sensors

Contact-free sensing based on WiFi channel state information (CSI) has shown considerable potential for human activity recognition and indoor localization. However, jointly addressing these two tasks remains challenging because raw CSI signals usually suffer from high-dimensional channel redundancy, task-irrelevant variations, and temporally entangled multi-scale fluctuations. To address these issues, this paper proposes a dual-task learning framework that emphasizes task-aligned subspace construction and structured temporal decomposition. Specifically, a Multi-task Aligned Re-ranking Subspace Principal Component Analysis (MARS-PCA) module is designed to re-rank principal components according to their discriminative relevance to both activity recognition and localization, thereby retaining a compact CSI representation that is more consistent with the dual-task objective. In addition, a multi-level wavelet decomposition front-end is introduced to separate CSI temporal responses into sub-band components, allowing transient activity-related dynamics and relatively stable location-related patterns to be represented more explicitly. The refined and decomposed features are then modeled by a lightweight temporal prediction module with channel-wise task regulation. Experiments on a public WiFi CSI dataset show that the proposed method achieves good performance in both activity recognition and indoor localization.

Towards high-dimensional IMU-based human activity recognition: data construction via 3D body modeling and classification with a multi-channel attention fusion network

Towards real-time sensor-based human activity recognition: a re-parameterized multidimensional feature communication fusion framework

The problem of Human Action Recognition (HAR) continues to be difficult with the intricate temporal interactions, superfluous frames, and minor visual variations that usually define similar actions. A large number of current approaches are based on either transformer or multimodal architectures, which are computationally costly and cannot be used in real-time or resource constrained systems. To overcome these shortcomings, we proposed a novel and lightweight HAR model that integrates spatial, temporal and consistency modeling in an efficient architecture. In our model, we combine EfficientNet-B2, used in efficient spatial feature extraction, with Unitary Temporal Encoder (UTE), to train long-range temporal dependencies, and Adaptive Temporal Consistency Module (ATCM), to improve local temporal consistency. The proposed system is trained and tested on the datasets UCF101 and HMDB51 with a Top-1 accuracy of 97.10% and 87% under RGB input only which is competitive among RGB-based HAR methods while maintaining low computational cost. The 9.3 million parameter model with an inference rate of 9.5 ms per 16-frame video clip is highly accurate and efficient, thus suitable in inference in real-time and at edges. Ablation studies further demonstrate that the proposed components contribute consistent performance improvements across the evaluated benchmarks.

Deep-learning-based human activity recognition (HAR) has been widely studied and applied in recent years, but it raises privacy concerns. Federated learning (FL) enables collaborative training without sharing raw data, thereby protecting user privacy. However, FL for HAR is challenged by three coupled factors: non-IID data across clients, aggregation under heterogeneous local models, and stringent computation and bandwidth budgets on edge hardware. To address these factors, this work introduces FedSynHAR, a lightweight FL framework that combines Gradient-Importance-based Adaptive Pruning (GIAP) with Channel-guided Feature-level Mutual Distillation (CFMD). GIAP prunes both server and client networks based on gradient importance, reducing computational and communication overhead; CFMD uses channel importance to guide mutual distillation between local and proxy models and mitigates pruning-induced degradation to improve robustness under non-IID conditions. Experiments on UCI-HAR and PAMAP2 demonstrate the effectiveness of FedSynHAR for federated HAR. On UCI-HAR, FedSynHAR converges about 2× faster than FedAvg, achieves 94.91% accuracy under non-IID settings, and reduces overhead by up to two orders of magnitude. Results on PAMAP2 further support the robustness of FedSynHAR under stronger heterogeneity.

EP-DConvFormer: an entropy-preserving and frequency-aware transformer for radar-based human activity recognition

Human Action Recognition (HAR) has become a field of great importance and interest for solving problems and accomplishing tasks across diverse areas. While the literature is vast, researchers face significant challenges in selecting optimal architectures, models and datasets that balance high accuracy with computational efficiency. To address this, the present systematic comprehensive review, conducted using the SALSA methodology, analyzes 81 articles and explores the best performing HAR methods. The review identified key benchmarks, notably NTU-RGB+D60 (15.15%) and UCF101 (12.12%). Most reviewed studies indicate that the most suitable models for these tasks are Convolutional Neural Networks (CNNs), which have demonstrated outstanding performance in analyzing graphical material, followed by Three-Dimensional Convolutional Neural Networks (3D CNNs) with temporal modeling (e.g., LSTMs) consistently yielding superior performance. Notably, advanced hybrid and skeleton-based models achieved peak accuracy rates of 98.3% and 98.7% on complex benchmarks, significantly outperforming traditional approaches.

Human action recognition (HAR) in the workshop-style environment poses unique challenges due to class imbalance, fused action boundaries, and context dependence. This study proposes a multi-stage fine-tuning supervised framework for workshop-style HAR based on a 3D CNN, supplemented by an unsupervised post-assessment. A total of 208 clips (100 frames per clip) from two internal and two external datasets were used, along with 183 clips generated from 152 videos in a subset of the public UCF101 dataset for an add-on supplementary analysis. An I3D-ResNet50 (I3D_R50) backbone was used to examine a pilot model for domain adaptation and subsequently optimized for cross-context generalization. In Study 1, the I3D_R50 model with a customized 57-class prediction head was trained on the internal dataset, achieving a solid class- and cluster-level performance, however, spurious co-activations and limited separation among similar actions were observed. Study 2 refined sample allocation and data augmentation, and applied threshold tuning to evaluate within-context adaptation (Test 1) and cross-context generalization (Test 2). Multi-level analyses (class-, clip-, cluster-, and person-level) revealed moderate co-activation and semantic fusion within context, whereas cross-context generalization suffered from structure over-merging and semantic drift. As a post-assessment, Study 3 provided supplementary evidence that the model captured certain semantic co-occurrence relationships, although its performance remained constrained by limited IGE-TREE specific training samples and fine-grained behavioral categories. Limitations include insufficient samples, blurred temporal sequence, and context dependency. Future work could explore domain-adversarial alignment, pose-guided normalization, and temporal contrastive modeling, and selective merging of semantically similar actions to enhance semantic disentanglement and generalization.

Wi-Fi sensing provides a privacy-preserving and device-free sensing modality for stationary crowd counting with a low deployment cost. However, labeled channel state information (CSI) data are difficult to obtain at scale, and CSI distributions vary significantly across deployment environments, leading to limited generalization. While semisupervised learning (SSL) has shown promise in Wi-Fi-sensing tasks, existing approaches primarily focus on human activity recognition (HAR) or localization and do not effectively address pseudolabel reliability or cross-domain robustness in crowd counting. To address these challenges, we propose a Wi-Fi-based crowd counting via domain-adversarial semisupervised learning (WiCount-DASL), a domain-adversarial SSL framework that jointly leverages limited labeled data and abundant unlabeled data while aligning feature representations across scenarios. The framework incorporates a classwise, adaptive pseudolabel thresholding mechanism and a targeted signal-level augmentation strategy to improve pseudolabel quality and robustness. Extensive experiments across multiple real-world deployment scenarios demonstrate that WiCount-DASL achieves competitive and robust counting accuracy under limited labeled data compared with representative baselines.

In this paper, we tackle a core challenge for wearable human activity recognition (HAR), namely the recognition of daily-living and locomotion activities from inertial sensor windows: delivering reliable, interpretable uncertainty on tiny microcontrollers where latency, RAM, and energy are tightly constrained. Existing embedded approaches either calibrate softmax confidences, which are cheap but brittle under sensor placement or tempo shifts, or rely on generative or posterior-sampling schemes that exceed TinyML budgets. We propose hyperdimensional distance- and uncertainty-aware human activity recognition (HDUQ-HAR), an on-device hyperdimensional computing (HDC) framework that encodes each IMU window into a bipolar hypervector, classifies via prototype similarity, and derives three complementary, lightweight uncertainty signals from the same representation: (1) distance to the class prototype, (2) similarity gap to the runner-up, and (3) vote-dispersion capturing n-gram consensus. A label-conditional conformal layer converts these scores into set-valued predictions with finite-sample coverage guarantees and exposes a human-readable reason code indicating why uncertainty increased. Across UCI HAR, WISDM, PAMAP2, and OPPORTUNITY with subject-out splits and realistic shifts (orientation, gain, time-warp, missing axis, cross-placement), HDUQ-HAR achieves near-target coverage at [Formula: see text] with near-singleton sets on i.i.d. data (average size 1.18-1.25) and robust shift/OOD detection (AUROC 0.92-0.96), while running in 3-5ms/window on Cortex-M4 with ∼6-9KB RAM and ∼5-7KB Flash. By unifying HDC geometry with label-conditional conformal prediction, our method shows that efficiency and reliability can co-exist in wearables, yielding small, calibrated sets that expand gracefully under shift and actionable explanations practitioners can trust.

Human Activity Recognition (HAR) stands as a crucial technology, with applications ranging from healthcare monitoring to sports analytics. However, the traditional approach to HAR is often time-consuming and susceptible to human errors due to the high complexities involved in processing diverse sensor data. Recognizing the imperative for efficiency and accuracy in HAR systems, we propose the development of an Automated Decision-maker (ADM) system. This system serves to automate HAR pipelines, addressing the challenges posed by the huge sensor data. By harnessing the power of automation, ADM significantly streamlines the HAR process, reducing the time required for hyperparameter tuning and minimizing the risk of human errors. The results obtained from our proposed ADM system demonstrate notable improvements in HAR performance, showcasing achieved accuracy of 96.436% for UCI-HAR & 99.783% for PAMAP2 datasets. Moreover, ADM can be described as an innovative approach that contributes to the optimization of HAR systems while also establishing a foundation for building robust and reliable systems in complex environments.

Role of prior in human activity recognition: A survey

DPBGFL: Debiased prototype bidirectional guided federated learning for human activity recognition

Activity-dependent resolution adjustment for radar-based human activity recognition

An efficient human action recognition framework for high-concurrency worker safety monitoring

Unified adaptive open-set incremental learning for industrial human action recognition

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.

Falls are a major contributor to morbidity and mortality among older adults, and timely fall detection can help reduce the severity of fall-related outcomes. Wearable inertial measurement unit (IMU) sensors offer a promising solution for fall detection; however, many existing approaches rely on multiple sensing locations and computationally intensive models, which can limit their practicality for resource-constrained wearable devices. This study proposes a knowledge distillation framework for efficient wrist-based fall detection using the publicly available University of Málaga fall detection dataset (UMAFall), a benchmark dataset for human activity recognition and fall detection. Although UMAFall was not collected from older adults, it provides a useful public benchmark for evaluating IMU-based fall detection methods. Knowledge distillation was implemented using a teacher–student framework, in which a high-capacity teacher model trained with IMU data from four body locations (waist, wrist, ankle, and chest) provided soft targets for guiding a compact wrist-only CNN student model. In a held-out test evaluation using Subjects 2 and 5, the teacher model achieved 97.6% accuracy and an F1 score of 96.7%, with approximately 1.3 million trainable parameters. The independently trained wrist-based CNN achieved 90.2% accuracy and an F1 score of 87.1%. After applying knowledge distillation, the student model improved to 95.1% accuracy and an F1 score of 93.3% while maintaining the same lightweight architecture. A supplementary leave-one-subject-out analysis showed slightly higher and more stable AUC for KD-CNN than the independently trained CNN (0.96 ± 0.03 vs. 0.94 ± 0.07). These findings suggest that knowledge distillation can improve wrist-only fall detection in this feasibility evaluation, but further validation using older adults and real-world smartwatch data is needed.

Recognizing human actions from static images in complex industrial environments remains challenging due to insufficient feature representation and high computational complexity. This issue is particularly critical in power-grid safety monitoring, where improper worker postures (e.g., bending, climbing, falling) can lead to severe accidents and personal injuries, necessitating automated monitoring systems that operate reliably on resource-constrained edge devices. This study proposes a bio-inspired lightweight recognition framework that integrates an improved YOLO-Pose model with a gated recurrent unit (GRU) network. The scientific motivation is grounded in the observation that the human musculoskeletal system achieves highly efficient motion perception through three key mechanisms: hierarchical muscle coordination providing intrinsic rotation invariance, proprioceptive feedback enabling real-time error correction, and selective neural gating reducing redundant information transmission. These biological principles directly inspire our technical contributions: polar-coordinate encoding provides rotation invariance, three-stage filtering mimics proprioceptive feedback, and GRU gating mirrors selective information propagation. Unlike prior approaches that treat pose-based action recognition as a generic computer vision problem, this work explicitly incorporates anatomical structural constraints into the computational pipeline. The framework addresses three research gaps: (1) existing methods lack biomechanically derived invariance properties; (2) GCN-based approaches use fixed topologies that fail to adapt to occlusion patterns; (3) the trade-off between model complexity and accuracy remains unsatisfactory for edge deployment. Experiments on the self-constructed SKPose dataset demonstrate that the proposed method achieves 95.04% accuracy, outperforming ST-GCN by 3.67 percentage points and 2s-AGCN by 1.94 percentage points, with an inference speed of 48 FPS on 8.7 M parameters in underground power-grid environments and provides practical support for biomimetic perception systems and industrial safety monitoring.