Hierarchical Human Activity Recognition Based on Smartwatch Sensors Using Branch Convolutional Neural Networks
5
Issued Date
2022-01-01
Resource Type
ISSN
03029743
eISSN
16113349
Scopus ID
2-s2.0-85142673585
Journal Title
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume
13651 LNAI
Start Page
52
End Page
60
Rights Holder(s)
SCOPUS
Bibliographic Citation
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) Vol.13651 LNAI (2022) , 52-60
Suggested Citation
Hnoohom N., Maitrichit N., Mekruksavanich S., Jitpattanakul A. Hierarchical Human Activity Recognition Based on Smartwatch Sensors Using Branch Convolutional Neural Networks. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) Vol.13651 LNAI (2022) , 52-60. 60. doi:10.1007/978-3-031-20992-5_5 Retrieved from: https://repository.li.mahidol.ac.th/handle/123456789/84332
Title
Hierarchical Human Activity Recognition Based on Smartwatch Sensors Using Branch Convolutional Neural Networks
Author's Affiliation
Other Contributor(s)
Abstract
Human activity recognition (HAR) has become a popular research topic in artificial intelligence thanks to the development of smart wearable devices. The main goal of human activity recognition is to efficiently recognize human behavior based on available data sources such as videos and images, including sensory data from wearable devices. Recently, HAR research has achieved promising results using learning-based approaches, especially deep learning methods. However, the need for high performance is still an open problem for researchers proposing new methods. In this work, we investigated the improvement of HAR by hierarchical classification based on smartwatch sensors using deep learning (DL) methods. To achieve the research goal, we introduced branch convolutional neural networks (B-CNNs) to accurately recognize human activities hierarchically and compared them with baseline models. To evaluate the deep learning models, we used a complex HAR benchmark dataset called WISDM-HARB dataset that collects smartwatch sensor data from 18 physical activities. The experimental results showed that the B-CNNs outperformed the baseline convolutional neural network (CNN) models when the hierarchical connection between classes was not considered. Moreover, the results confirmed that branch CNNs with class hierarchy improved the recognition performance with the highest accuracy of 95.84%.