Tuesday, November 17, 2026
Science
No Result
View All Result
  • Login
  • HOME
  • SCIENCE NEWS
  • CONTACT US
  • HOME
  • SCIENCE NEWS
  • CONTACT US
No Result
View All Result
Scienmag
No Result
View All Result
Home Science News Technology and Engineering

CoHAR Framework Improves Wearable Action Recognition Accuracy Across Users and Sensor Placements

November 17, 2026
in Technology and Engineering
Denise Maddox
By Denise Maddox Scienmag Editorial Profile - Mechanical Engineering
Reading Time: 6 mins read
0
CoHAR Framework Improves Wearable Action Recognition Accuracy Across Users and Sensor Placements

CoHAR Framework Improves Wearable Action Recognition Accuracy Across Users and Sensor Placements

65
SHARES
587
VIEWS
Share on FacebookShare on Twitter
ADVERTISEMENT

Computer scientists have designed an unsupervised cross-domain adaptation framework named CoHAR to address one of the most persistent bottlenecks in wearable-based human action recognition: the severe degradation of model performance when algorithms encounter new individuals, distinct athletic proficiencies, and altered sensor body placements. While wearable devices such as smartphones and smartwatches now routinely incorporate inertial, gyroscopic, and magnetometric sensors to monitor movements in daily life, healthcare, and sports analytics, conventional deep learning models depend heavily on supervised training regimes that demand extensive, expert-annotated datasets. In real-world environments, collecting manual labels for every new user or activity variant is practically infeasible. Furthermore, human movements exhibit substantial cross-user variations stemming from individual body characteristics, physical coordination, gender differences, and evolving motor skill levels. By pairing source-labeled data with completely unlabeled target user data, the new architecture enables models to generalize across distinct users without requiring new manual annotations.

The underlying problem stems from the wide statistical discrepancies observed across different users performing identical physical actions. In their empirical analysis of accelerometer signals from badminton players performing micro-complex actions such as smashes and volleys, the investigators demonstrated that movement magnitudes differ dramatically by user skill level and biological sex. For example, experienced players produced distinct, highly coordinated acceleration patterns with magnitude fluctuations spanning from 0.52 to 0.98, whereas intermediate players recorded values between 0.18 and 0.88, and novice players registered a much narrower range between 0.48 and 0.54. Systematic discrepancies also appeared between genders, where male players exhibited acceleration magnitudes spanning from 0.12 to 0.97, compared to female players who exhibited a narrower band between 0.33 and 0.48. In addition, shifting sensors across the body induced massive domain shifts: signals captured from the hand and left leg reached higher magnitudes of 0.6 to 1.0 and 0.6 to 0.9, respectively, whereas sensors located on the palm and right leg registered much lower magnitude ranges of 0.3 to 0.7 and 0.1 to 0.5. These substantial distribution shifts severely compromise the predictive accuracy of standard classifiers.

To overcome these distribution shifts without target labels, the researchers formulated CoHAR around a single-stage joint optimization framework composed of complementary functional modules. The baseline classification model begins with a supervised deep convolutional feature extractor trained on labeled source domain data. This neural network architecture incorporates three convolutional layers structured with an increasing number of filters and progressively smaller kernel sizes, with each convolutional block paired with batch normalization and max-pooling operations to mitigate internal covariate shifts and extract both high-level and low-level temporal features. The network terminates in two fully connected layers followed by a softmax classification head that minimizes standard cross-entropy loss over the available source instances. However, because new users provide only unlabeled sensor data, the classifier cannot directly align target features using conventional supervised loss functions, requiring specialized adaptation mechanisms to bridge the domain gap.

A critical component in CoHAR is the generation of dependable pseudo-labels for unlabeled target data through temporal ensembling. In traditional unsupervised domain adaptation, models often employ self-labeling or vanilla pseudo-labeling, but erroneous initial classifications inevitably propagate through subsequent iterations, leading to model degradation and training collapse. To resolve this instability, CoHAR implements an exponential moving average update that aggregates model predictions across previous training iterations using an explicit temporal decay factor alpha. By maintaining an ensemble of historical predictions, the framework smooths out stochastic fluctuations and filters out noisy pseudo-label assignments. The mathematical formulation explicitly balances historical predictions against current iteration outputs, allowing the model to stabilize learning even in the presence of underrepresented minority activity classes. Systematic grid searches revealed that setting the decay factor alpha within a range of 0.55 to 0.75 produces optimal stability, with an alpha parameter of 0.60 yielding the fastest training convergence and achieving pseudo-label correctness of approximately 72 percent on held-out evaluation samples.

To eliminate conditional distribution discrepancies between source and target feature spaces, the framework introduces a class-wise conditional maximum mean discrepancy metric evaluated in a Reproducing Kernel Hilbert Space, termed kCMMD. Rather than merely aligning global marginal feature distributions regardless of activity class, kCMMD maps source and pseudo-labeled target feature representations into a high-dimensional space using a Radial Basis Function kernel to align conditional distributions class by class. The objective function calculates intra-domain similarities within source samples and within target samples, and subtracts twice the inter-domain similarity across both domains, establishing a symmetrical adaptation step that pulls matching class representations together. Furthermore, the researchers introduced a regularization identity matrix scaled by a hyperparameter lambda added along the diagonals of the source and target kernel matrices. This mathematical term prevents matrix ill-conditioning and controls overfitting, ensuring that the learned latent representations remain stable and well-conditioned across iterative updates.

To fortify the model against subtle physical noise and spatial orientation variations, the researchers combined consistency regularization with unsupervised geometric data augmentations. The augmentation pipeline applies two targeted transformations specifically calibrated for wearable sensor time series: stochastic jitter and rotational displacement. Jitter introduces high-frequency Gaussian noise with a standard deviation between 0.01 and 0.10, adding controlled stochastic variability that simulates minor sensor noise and subtle orientation tremors without corrupting underlying motion patterns; optimal empirical stability was observed with jitter values between 0.03 and 0.06. Simultaneously, spatial rotation within an angular displacement interval of minus pi over four to positive pi over four alters coordinate orientation, reflecting physical shifts in sensor placement. Systematic cosine similarity evaluations between unaugmented and rotated samples confirmed that similarity scores never dropped below 0.80 across the entire range, remaining above 0.95 for rotations within plus or minus thirty degrees. Consistency regularization then minimizes the Kullback-Leibler divergence between predictions generated from original sensor inputs and their augmented counterparts, ensuring that the model maintains consistent probabilistic predictions across perturbations.

The complete training pipeline integrates the source cross-entropy loss, the kCMMD conditional domain adaptation loss, and the dual-domain consistency regularization loss into a single unified objective function. Hyperparameter weights beta zero and beta one scale the adaptation and consistency terms relative to the supervised classification loss, both set to numerical values of one to maintain balanced gradient contributions. During training experiments conducted on a Linux server equipped with Intel processors and four dedicated graphics processing units using PyTorch, the total training loss demonstrated steady numerical convergence, stabilizing fully within 128 training epochs. Extensive quantitative evaluations across five benchmark datasets—including the Badminton Activity Recognition dataset, Daily and Sports Activity dataset, Physical Activity Monitoring dataset, Workout Activities dataset, and the MPSC-AR dataset—demonstrated that CoHAR consistently outpaced six state-of-the-art unsupervised domain adaptation methods, yielding average macro-F1 score improvements ranging between 4 and 12 percent across five independent trials per dataset.

Rigorous evaluations across specific domain shift scenarios highlighted the robustness of the integrated architecture. In cross-gender adaptation experiments where models were trained exclusively on male subjects and evaluated on female subjects, or trained on female subjects and tested on male subjects, CoHAR surpassed existing baselines with an average macro-F1 score improvement of 5.3 plus or minus 2.14 percent. In cross-dexterity experiments tracking transitions from expert to beginner gym participants on the MMDOS dataset, CoHAR outscored the leading baseline, SWL-Adapt, by approximately 11 percent. Furthermore, across cross-body sensor adaptation scenarios where algorithms trained on upper-body wearable sensors were deployed on lower-body sensor streams, CoHAR consistently outperformed competing frameworks with average macro-F1 gains of 1 to 5 percent. Statistical significance testing using non-parametric analysis confirmed that these performance gains were statistically dependable, yielding a p-value of 0.02471.

Qualitative investigations and metric analyses further validated the underlying representation mechanics of the system. Two-dimensional t-SNE visualizations of learned feature embeddings on target users showed tight clustering of inter-class activities alongside distinctly separated intra-class decision boundaries, demonstrating clear preservation of action semantics even under strong user shifts. To complement this visualization, the researchers computed the Proxy A-distance metric to directly measure domain discrepancy between source and target feature spaces. Following adaptation with CoHAR, the Proxy A-distance dropped markedly from 1.79 plus or minus 0.03 down to 0.84 plus or minus 0.02, confirming a substantial reduction in domain distribution divergence. Ablation experiments isolating individual modules revealed that combining temporal ensembling, kCMMD loss, and consistency regularization drove cumulative performance gains between 22 and 35 percent over unadapted baselines across distinct augmentation configurations.

Despite these demonstrable gains, the authors identified several structural limitations that warrant future research and refinement. The framework currently relies on predefined geometric and frequency perturbations, which lack input-aware adaptability and may not capture all out-of-distribution variations encountered in unstructured daily environments. Additionally, computing class-wise kernel matrices over mini-batches introduces a worst-case computational time complexity of big-O of d times B squared and memory complexity of big-O of B squared, where B is the combined batch size and d is the feature dimension, potentially straining resource-constrained edge hardware. The authors noted that comprehensive reporting of on-device training latency, inference time, and memory footprints remains necessary before deploying CoHAR to low-power microcontrollers. Future work is planned to incorporate multimodal sensor fusion, input-aware adaptive augmentations, cross-covariance alignment methods, open-set adaptation for unrecognized activity categories, and post-calibration mechanisms to further refine prediction uncertainty under extreme distribution shifts.

Subject of Research: Unsupervised domain adaptation for wearable human activity recognition

Article Title: Unsupervised cross-domain adaptation for wearable-based human activity recognition

Article References: Ghosh, I., Chugh, G., Faridee, A. Z. M., & Roy, N. (2026). Unsupervised cross-domain adaptation for wearable-based human activity recognition. Knowledge and Information Systems, 68(1), Article 318. https://doi.org/10.1007/s10115-026-02918-8

Image Credits: AI Generated

DOI: 10.1007/s10115-026-02918-8

Keywords: wearable sensors, human activity recognition, unsupervised domain adaptation, pseudo-labeling, temporal ensembling, maximum mean discrepancy, consistency regularization, deep learning, accelerometer data, cross-user adaptation, sensor placement, pattern recognition, inertial measurement units, data augmentation, motion analysis, biometrics

Cite Scienmag News

Denise Maddox. (November 17, 2026). CoHAR Framework Improves Wearable Action Recognition Accuracy Across Users and Sensor Placements. Scienmag. https://scienmag.com/cohar-framework-improves-wearable-action-recognition-accuracy-across-users-and-sensor-placements/

Denise Maddox. "CoHAR Framework Improves Wearable Action Recognition Accuracy Across Users and Sensor Placements." Scienmag, 17 November 2026, https://scienmag.com/cohar-framework-improves-wearable-action-recognition-accuracy-across-users-and-sensor-placements/. Accessed 17 November 2026.

Denise Maddox. "CoHAR Framework Improves Wearable Action Recognition Accuracy Across Users and Sensor Placements." Scienmag. November 17, 2026. https://scienmag.com/cohar-framework-improves-wearable-action-recognition-accuracy-across-users-and-sensor-placements/

Tags: accelerometer dataBiometricsConsistency regularizationcross-user adaptationdata augmentationdeep learninghuman activity recognitioninertial measurement unitsmaximum mean discrepancymotion analysisPattern Recognitionpseudo-labelingsensor placementtemporal ensemblingunsupervised domain adaptationwearable sensors
Share26Tweet16
Previous Post

Higher Temperatures Shorten Extrinsic Incubation Periods for Select Mosquito Viruses

Next Post

Historic Drainage Rules Hamper Wetland Restoration on Swedish Island of Öland

Related Posts

Solvothermal Synthesis Yields Bismuth Selenide Nanoparticles with Distinct Phonon Modes
Technology and Engineering

Solvothermal Synthesis Yields Bismuth Selenide Nanoparticles with Distinct Phonon Modes

November 17, 2026
Simple Routing Rules Beat Complex AI for Manufacturing Under Machine Breakdowns
Technology and Engineering

Simple Routing Rules Beat Complex AI for Manufacturing Under Machine Breakdowns

November 17, 2026
Neuromorphic IoT Hardware Cuts Edge Energy Consumption by 89.7 Percent
Technology and Engineering

Neuromorphic IoT Hardware Cuts Edge Energy Consumption by 89.7 Percent

November 17, 2026
Prestressed Reaction Wall System Retains Stable Compression Across Construction and Loading
Technology and Engineering

Prestressed Reaction Wall System Retains Stable Compression Across Construction and Loading

November 17, 2026
Full-Sized Humanoid Platform HART Navigates Vehicles Using Landmark Wayfinding Architecture
Technology and Engineering

Full-Sized Humanoid Platform HART Navigates Vehicles Using Landmark Wayfinding Architecture

November 17, 2026
New Stochastic Path Sampler Improves Efficiency in Lattice Field Theory
Technology and Engineering

New Stochastic Path Sampler Improves Efficiency in Lattice Field Theory

November 17, 2026
Next Post
Historic Drainage Rules Hamper Wetland Restoration on Swedish Island of Öland

Historic Drainage Rules Hamper Wetland Restoration on Swedish Island of Öland

  • Mothers who receive childcare support from maternal grandparents show more optimized

    Mothers who receive childcare support from maternal grandparents show more parental warmth, finds NTU Singapore study

    27656 shares
    Share 11059 Tweet 6912
  • University of Seville Breaks 120-Year-Old Mystery, Revises a Key Einstein Concept

    1061 shares
    Share 424 Tweet 265
  • Bee body mass, pathogens and local climate influence heat tolerance

    682 shares
    Share 273 Tweet 171
  • Researchers record first-ever images and data of a shark experiencing a boat strike

    546 shares
    Share 218 Tweet 137
  • Groundbreaking Clinical Trial Reveals Lubiprostone Enhances Kidney Function

    531 shares
    Share 212 Tweet 133
Science

Embark on a thrilling journey of discovery with Scienmag.com—your ultimate source for cutting-edge breakthroughs. Immerse yourself in a world where curiosity knows no limits and tomorrow’s possibilities become today’s reality!

RECENT NEWS

  • Digital Tax Adoption, IPSAS, and Tax Compliance Among Jordanian SMEs
  • Chinese Visceral Adiposity Index Shows Nonlinear Link to Cardiovascular Disease Risk
  • Review Links Skin Microbiota and T-Cell Crosstalk to Venous Disease Progression
  • Mathematical Fixes in Climate Model Ocean Flux Equations Resolve Simulation Errors

Categories

  • Agriculture
  • Anthropology
  • Archaeology
  • Athmospheric
  • Biology
  • Biotechnology
  • Blog
  • Bussines
  • Cancer
  • Chemistry
  • Climate
  • Earth Science
  • Editorial Policy
  • Marine
  • Mathematics
  • Medicine
  • Pediatry
  • Policy
  • Psychology & Psychiatry
  • Science Education
  • Science News
  • Social Science
  • Space
  • Technology and Engineering

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 5,151 other subscribers

© 2026 Scienmag - Science Magazine

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • HOME
  • SCIENCE NEWS
  • CONTACT US

© 2026 Scienmag - Science Magazine

Discover more from Science

Subscribe now to keep reading and get access to the full archive.

Continue reading