Zahin, Abrar and Tan, Le and Hu, Rose (2019) Sensor-Based Human Activity Recognition for Smart Healthcare: A Semi-supervised Machine Learning. In: Artificial Intelligence for Communications and Networks. First EAI International Conference, AICON 2019, Harbin, China, May 25–26, 2019, Proceedings, Part I.
32736.pdf
Download (1MB)
Abstract
Human action recognition is an integral part of smart health monitoring, where intelligence behind the services is obtained and improves through sensor information. It poses tremendous challenges due to huge diversities of human actions and also a large variation in how a particular action can be performed. This problem has been intensified more with the emergence of Internet of Things (IoT), which has resulted in larger datasets acquired by a massive number of sensors. The big data based machine learning is the best candidate to deal with this grand challenge. However, one of the biggest challenges in using large datasets in machine learning is to label sufficient data to train a model accurately. Instead of using expensive supervised learning, we propose a semi-supervised classifier for time-series data. The proposed framework is the joint design of variational auto-encoder (VAE) and convolutional neural network (CNN). In particular, the VAE intends to extract the salient characteristics of human activity data and to provide the useful criteria for the compressed sensing reconstruction, while the CNN aims for extracting the discriminative features and for producing the low-dimension latent codes. Given a combination of labeled and raw time-series data, our architecture utilizes compressed samples from the latent vector in a deconvolutional decoder to reconstruct the input time-series. We intend to train the classifier to detect human actions for smart health systems.
| Item Type: | Conference or Workshop Item (UNSPECIFIED) |
|---|---|
| Date Deposited: | 04 Mar 2026 12:14 |
| Last Modified: | 20 Nov 2026 11:03 |
| URI: | https://eprints.eai.eu/id/eprint/23486 |
