A context-token transformer framework for multi-resident smart home human activity recognition with comparative evaluation against recurrent architectures

Human Activity Recognition has emerged as a key enabling technology for context-centric intelligent home environments, facilitating applications such as ambient intelligence, healthcare monitoring, assisted living, and elderly residents monitoring. Despite substantial advancements in conventional Human Activity Recognition strategies as well as intelligent learning-based Human Activity Recognition mechanisms, reliable and robust activity recognition in real-world intelligent home environments remains challenging due to contextual ambiguity, overlapping activity patterns, and significant inter- and intra-subject variability, typically in multi-resident situations. In this paper, we present a comprehensive investigation of a unified context-aware Human Activity Recognition architecture for intelligent smart home environments, explicitly evaluated both single-resident and multi-resident activity recognition. In this work, four representative Deep Learning architectures, specifically Bidirectional Long Short-Term Memory with Attention, Cascading Long Short-Term Memory, TransformerHAR, and the Patch and Context-Token-Transformer-based framework, were comprehensively investigated across single-resident and multi-resident smart home environments to measure their effectiveness under varying levels of activity complexity and resident interactions. The extensive experiments, validated through cross-validation and repeated-trial robustness analysis, are conducted using MARBLE and CASAS-Kyoto datasets to investigate performance under both single-resident and multi-resident smart home environments. These experimental results indicate that the Patch and Context Based Token Transformer framework accomplishes competitive performance in single-resident multiple activity recognition settings while significantly enhancing activity recognition performance in multi-resident settings, attaining an average accuracy of 69.38% compared with 58.66% (Bidirectional Long Short-Term Memory with Attention), 58.96% (TransformerHAR), and 60.41% (Cascading Long Short-Term Memory with Attention) for these competing baselines. Further, comprehensive class/activity-wise, confusion-matrix, and convergence investigations validate that the context-token conditioning competently assists in improving activity disambiguation, learning/training stability, and generalization, typically in multi-resident scenarios characterized by overlapping sensor patterns, concurrent activities, and complex contextual interactions.

Authors

Institutions

Publication Details

Journal
Scientific Reports
Published
2026-11-13
DOI
https://doi.org/10.1038/s41598-026-75805-5
Primary Topic
Context-Aware Activity Recognition Systems
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

A context-token transformer framework for multi-resident smart home human activity recognition with comparative evaluation against recurrent architectures

Thinagaran Perumal, Sulabh Bansal, Abhishek Sharma, Varun Kumar Sharma et al.
Scientific Reports
Context-Aware Activity Recognition Systems
article

A context-token transformer framework for multi-resident smart home human activity recognition with comparative evaluation against recurrent architectures

Thinagaran Perumal, Sulabh Bansal, Abhishek Sharma, Varun Kumar Sharma, Ujjwal Krishna Khanna, Pratham Kala
article en

Abstract

Human Activity Recognition has emerged as a key enabling technology for context-centric intelligent home environments, facilitating applications such as ambient intelligence, healthcare monitoring, assisted living, and elderly residents monitoring. Despite substantial advancements in conventional Human Activity Recognition strategies as well as intelligent learning-based Human Activity Recognition mechanisms, reliable and robust activity recognition in real-world intelligent home environments remains challenging due to contextual ambiguity, overlapping activity patterns, and significant inter- and intra-subject variability, typically in multi-resident situations. In this paper, we present a comprehensive investigation of a unified context-aware Human Activity Recognition architecture for intelligent smart home environments, explicitly evaluated both single-resident and multi-resident activity recognition. In this work, four representative Deep Learning architectures, specifically Bidirectional Long Short-Term Memory with Attention, Cascading Long Short-Term Memory, TransformerHAR, and the Patch and Context-Token-Transformer-based framework, were comprehensively investigated across single-resident and multi-resident smart home environments to measure their effectiveness under varying levels of activity complexity and resident interactions. The extensive experiments, validated through cross-validation and repeated-trial robustness analysis, are conducted using MARBLE and CASAS-Kyoto datasets to investigate performance under both single-resident and multi-resident smart home environments. These experimental results indicate that the Patch and Context Based Token Transformer framework accomplishes competitive performance in single-resident multiple activity recognition settings while significantly enhancing activity recognition performance in multi-resident settings, attaining an average accuracy of 69.38% compared with 58.66% (Bidirectional Long Short-Term Memory with Attention), 58.96% (TransformerHAR), and 60.41% (Cascading Long Short-Term Memory with Attention) for these competing baselines. Further, comprehensive class/activity-wise, confusion-matrix, and convergence investigations validate that the context-token conditioning competently assists in improving activity disambiguation, learning/training stability, and generalization, typically in multi-resident scenarios characterized by overlapping sensor patterns, concurrent activities, and complex contextual interactions.

Scientific Reports
Universiti Putra Malaysia (MY), LNM Institute of Information Technology (IN), Manipal University Jaipur (IN)
Context-Aware Activity Recognition Systems
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon