Human Activity Recognition (HAR) is the automated process of identifying human actions using sensor data or video, which is widely used in healthcare, smart environments, and surveillance. Although HAR based on computer vision has advanced rapidly, existing reviews do not adequately address the recent shift toward hybrid deep-learning architectures or provide a structured comparison of the trade-offs relevant to real-world deployment. This literature review addresses that gap through a PRISMA-guided analysis of articles published between 2021 and 2025 and retrieved from four major databases. The review develops a reproducible taxonomy of nine architectural families and applies a multidimensional evaluation framework covering classification accuracy, computational efficiency for edge deployment, environmental generalization, and fine-grained activity recognition. The findings show that hybrid architectures are the dominant design strategy, while attention-based and graph-based models play important specialized roles depending on temporal complexity, privacy requirements, and deployment constraints, with the literature concentrated mainly in healthcare and security applications.
Human Activity Recognition (HAR) is a fast-growing research area that focuses on identifying human actions using data collected from sensors and vision-based devices. It plays an important role in applications like health monitoring, smart homes, surveillance, sports analysis, and human-computer interaction. In recent years, several methods have been developed to improve the performance of HAR systems using machine learning, deep learning, and hybrid models. This paper presents a detailed review of different methods used in HAR. The study is divided into three main categories: vision-based methods, sensor-based methods, and hybrid approaches that combine both types. Each method is discussed with examples from recent research, along with their advantages and limitations. A comparison is also provided in the form of a table to highlight the performance and challenges of each approach. Although HAR systems have achieved good results in controlled environments, several challenges still remain. These include poor generalization to new users or unknown environments, difficulty in recognizing complex or overlapping activities, dependence on large datasets, and lack of real-time performance. This paper also discusses these research gaps based on recent findings. The future of HAR depends on building more accurate, reliable, and real-time systems that can adapt to different situations. The paper concludes by suggesting possible directions for future work, such as the development of lightweight models, use of standard datasets, better handling of real-time data, and making models more interpretable.
Human Activity Recognition (HAR) is increasingly being incorporated in intelligent surveillance systems; however, the majority of the current techniques do not perform well in real-world situations that incorporate live video streams, dynamic backgrounds and complex human behaviours. In particular, there is a significant difference in comparing HAR algorithms when they operate on trimmed (segmented) video vs. untrimmed (streaming) video. This article provides an overview of vision-based HAR systems designed for surveillance applications. The article compares the performance of traditional machine learning techniques with recent deep learning architectures such as CNNs, RNNs, TCNs, and Transformers, based on accuracy, computational cost, and suitability for real-world disorderly plotting. The article also examines how well publication datasets model video surveillance scenarios and describes the limitations of each dataset. The article identifies several practical issues associated with each technique and dataset (occlusions, illumination changes, camera motions, and crowded environments). This article also highlights the major contributions regarding the trimmed and untrimmed video-based HAR methodologies and provides an exhaustive review of models, datasets, and practical issues related to their real-world application within the context of surveillance. Finally, the authors also provide guidance on potential future directions of research, such as Edge AI, product data integration & explainable AI, and techniques focused on ensuring privacy.
Disha Deotale, M. Verma, P. Suresh et al.· Discover Artificial Intellig...· 0 citations
Human activity recognition (HAR) based on smartphone and wearable sensor data is commonly addressed using statistical learning methods and deep neural networks that often provide strong predictive performance, but at the expense of limited interpretability and substantial computational and energy requirements. Such limitations reduce their suitability for deployment in practical sensing environments where model decisions must be transparent, verifiable and executable on resource-constrained devices. In this work, we investigate the Convolutional Tsetlin Machine (CTM) for multimodal HAR using only the raw inertial signals (9 × 128) of the UCI-HAR dataset, rather than its pre-computed 561-feature representation. The Tsetlin Machine is a novel neuro-symbolic machine learning approach that offers two important advantages over many conventional machine learning methods: (i) it learns logic-based decision rules that support human inspection and provide a transparent basis for analyzing model decisions, and (ii) it operates with comparatively low computational complexity, making it well suited to efficient and low-power on-device learning. The proposed study systematically analyses the contribution of different feature modalities by decomposing the inertial signals space into semantically defined subsets according to: (i) sensor source: accelerometer and gyroscope; (ii) signal group: gyroscope angular velocity, body and total acceleration (including gravity); (iii) coordinate axis: x, y and z. A separate CTM classifier was trained for each modality and its combinations in order to determine the relative discriminative value of each modality group for activity classification. In addition to predictive performance, the study emphasizes the interpretability of the CTM model ensured by expressing each decision in the form of propositional clauses, thereby enabling visualization and direct inspection of the modality-specific patterns supporting each activity class. Owing to its symbolic structure and modest computational demands, the CTM provides a principled framework for the design of explainable, resource-efficient and deployable HAR systems. The proposed work therefore contributes toward trustworthy multimodal sensing by jointly addressing predictive performance, interpretability and suitability for embedded and mobile platforms.
O. Tarasyuk, A. Gorbenko, O. Gordieiev et al.· Italian National Conference...· 0 citations
Human Activity Recognition (HAR) from wearable sensors supports applications in healthcare, rehabilitation, fitness tracking, and smart environments. Yet, existing deep learning approaches require dataset-specific training, large labeled corpora, and repeated adaptation to new sensor settings or activity taxonomies. Retrieval-Augmented Generation for Human Activity Recognition (RAG-HAR) addresses this by framing HAR as a training-free, retrieval-augmented task, in which statistical descriptions of sensor windows are used to retrieve similar labeled examples that guide LLM-based classification. We introduce RAG-HAR+, a retrieval-first and cost-optimized extension that strengthens retrieval while reducing dependence on LLM-based inference. RAG-HAR+ uses an offline Retrieval Designer Agent to design dataset-specific feature groups from a diverse pool of motion descriptors, enabling sensor windows to be compared using features better aligned with dataset-specific activity patterns. During inference, RAG-HAR+ uses majority voting over retrieved neighbors for samples with strong retrieval evidence and defers only uncertain cases to an LLM-based Ambiguity Resolver Agent. Across six HAR benchmarks, RAG-HAR+ maintains competitive or improved performance while reducing LLM usage, token consumption, and inference time. We further extend the RAG-HAR mobile prototype to demonstrate the practical feasibility of retrieval-first, LLM-assisted HAR in mobile sensing scenarios.
Hansi Karunarathna, Nirhoshan Sivaroopan, Chamara Madarasingha et al.· 0 citations
WiFi-based human activity recognition (HAR) and fall detection have emerged as promising alternatives to vision- and wearable-based systems for modern assisted living and clinical monitoring for the elderly. These approaches exploit variations in Channel State Information, capturing amplitude and phase fluctuations across subcarriers to characterize how human motion perturbs these patterns through multipath propagation. This narrative review synthesizes 26 peer-reviewed studies published between 2017 and 2025, focusing on systems evaluated in real-world, non-laboratory environments. While recent deep learning models demonstrate near-perfect benchmark accuracy on fine-grained datasets, cross-environment assessments consistently reveal a severe “reality gap.” Specifically, several reviewed studies reported accuracy declines exceeding 20% when deployed in previously unseen buildings, while multi-user interference reduced recognition accuracy by 21–46%. Furthermore, dataset availability and reproducibility are critical considerations; only a limited number of reviewed studies utilized publicly available datasets, and none provided a standardized benchmarking guide-line. To address these gaps, this review presents a comprehensive taxonomy of post-2017 architectures, a comparative synthesis of laboratory versus deployment performance, and a critical analysis of key implementation challenges. We conclude by proposing a set of recommended reporting practices and outlining actionable future research. In doing so, the paper outlines key recommendations for developing adaptive, location-independent models, emphasizing the need for domain-transfer techniques, cross-building validation, and expanding open datasets to enhance generalizability.
Received: 31 August 2025 | Revised: 13 March 2026 | Accepted: 25 June 2026
Conflicts of Interest
The authors declare that they have no conflicts of interest to this work.
Data Availability Statement
Data sharing is not applicable to this article as no new data were created or analyzed in this study.
Author Contribution Statement
Kok Chung Chua: Conceptualization, Methodology, Validation, Formal analysis, Writing – original draft, Writing – review & editing, Visualization. Kai Liang Lew: Resources, Writing – original draft, Writing – review & editing, Project administration. Chean Khim Toa: Conceptualization, Validation, Resources, Writing – review & editing, Visualization, Supervision, Project administration, Funding acquisition. Matthew Alexander Paudianto: Investigation, Data Curation, Writing– original draft, Writing – review & editing. Ivan Nathanael: Investigation, Data curation, Writing – original draft, Writing – review & editing. Abel Nathanael Hutapea: Investigation, Data curation, Writing – original draft, Writing – review & editing.
Kok Chung Chua, Kai Liang Lew, Chean Khim Toa et al.· Artificial Intelligence and...· 0 citations
Falls represent a threat to older adults, overload healthcare systems, and reduce quality of life. Vision-based fall detection has advanced recently through deep learning, yet most proposed models lack validation on physical hardware and do not report inference-time metrics. This systematic review, following PRISMA and Kitchenham guidelines, targets this gap. We focus exclusively on vision-based systems that report inference speed on a specified device. We define real-time performance using a threshold of 10 fps, based on the reported duration of the critical fall phase in real-life falls. From 588 records across IEEE Xplore, ACM Digital Library, Web of Science Core Collection, and PubMed (2019–2024), only 11 met all inclusion criteria, highlighting how few studies validate real-time performance on physical hardware. The findings show that CNN-based architectures dominate algorithm choice, edge devices dominate deployment platforms, and optimization remains central to real-time inference on constrained hardware. Across these studies, we identify two persistent limitations: no real-world testing with older adults and reliance on small, controlled datasets with simulated falls.
Mahammad Nabizade, Réda Yahiaoui, Isabelle Lajoie et al.· Italian National Conference...· 0 citations