IoT-Based Real-Time Human Motion Recognition Based on Skeletons for Next Generation Healthcare using Deep Learning Techniques
Subrata Kumer Paul1, Rakhi Rani Paul2, Md. Ekramul Hamid3
1Subrata Kumer Paul, Department of CSE, University of Rajshahi, Rajshahi, Bangladesh, Department of CSE, Bangladesh Army University of Engineering & Technology (BAUET), Natore, Rajshahi, Bangladesh.
2Rakhi Rani Paul, Department of CSE, University of Rajshahi, Rajshahi, Bangladesh. Department of CSE, Bangladesh Army University of Engineering & Technology (BAUET), Natore, Rajshahi, Bangladesh.
3Ekramul Hamid, Department of CSE, University of Rajshahi, Rajshahi, Bangladesh.
Manuscript received on 08 January 2023 | Revised Manuscript received on 29 January 2023 | Manuscript Accepted on 15 February 2023 | Manuscript published on 28 February 2023 | PP: 33-46 | Volume-12 Issue-3, February 2023 | Retrieval Number: 100.1/ijitee.D105514040325 | DOI: 10.35940/ijitee.D1055.12030223
Open Access | Editorial and Publishing Policies | Cite | Indexing and Abstracting
© The Authors. Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP). This is an open access article under the CC-BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Abstract: The Internet of Things (IoT) and mobile technology have transformed healthcare by enabling real-time monitoring and diagnosis of patients. Human Motion Recognition (HMR) plays a vital role in healthcare systems in the 21st century, yet existing solutions face challenges such as high computational demands, low accuracy, and limited adaptability. This paper introduces a novel approach to real-time HMR using skeleton data with ENConvLSTM. This hybrid deep learning model combines EfficientNet and ConvLSTM to improve both accuracy and processing speed. The EfficientNet excels in image classification, while ConvLSTM processes sequential data, making this hybridization well-suited for dynamic motion analysis. Applied to the “NTU RGB+D 120” public dataset, our model achieves an accuracy of 94.85% for cross-subject and 96.45% for cross-view evaluations, addressing key healthcare-related activities across 12 medical classes. Moreover, the dataset presents several challenges, including a large data size, high dimensionality, class imbalance, noisy data, and limited domain labelling, all of which require advanced processing and modelling techniques. Our model’s unique use of skeleton data tracking 25 body joints enables the detection of subtle motion changes, ideal for early diagnosis and intervention in healthcare scenarios. Additionally, we integrate the system with a Raspberry Pi and a GSM module, providing real-time alerts via Twilio’s SMS service to instantly notify caregivers and patients. However, this solution is pivotal in modern healthcare, offering scalable, efficient, and proactive patient monitoring, leading to improved outcomes and reduced healthcare costs.
Keywords: Real-time Human Motion Recognition (HMR), EN Conv LSTM, Skeleton Data, NTU RGB+D 120 Dataset.
Scope of the Article: Computer Science and Applications