Energy-efficient human activity recognition framework for wearable devices : development, deployment, and analysis

Abstract

The rapid population aging has significant implications for society, especially in healthcare. At the same time, technological advancement and the spread of artificial intelligence provide new possibilities to meet these challenges and improve health and well-being, preventing a future collapse of the healthcare system. Wearable devices such as smart bands and smartwatches, already widely adopted today to monitor physical activities, will be important allies to promote better healthcare services. One area of research that combines these devices with machine learning is Human Activity Recognition (HAR), which has been widely discussed in the literature. The accelerated progress of deep learning techniques has also contributed to the development of models that surpassed the classification performance of the state-of-the-art until then. However, such models are highly complex to be deployed on resource-constrained wearable devices, which demand applications with low power consumption, real-time response, and privacy. Edge computing along with lighter and more efficient HAR systems is a viable solution to meeting these requirements. Given this scenario, this doctoral research aims to develop, implement and test an energy-efficient HAR framework for classifying human activities. The developed models were extensively analyzed in different configurations, and compared with reference and state-of-the-art models. In addition, the proposed framework was deployed and tested on a microcontroller, providing important analysis regarding computational performance. With the MHEALTH dataset, using only data sampled at 10 Hz from a wrist-worn sensor, the proposed framework achieved accuracy of 79.07%, surpassing reference classifiers, with a model of 44.58 kB, which consumes only 1.59 mW. In the multi-sensors scenario, the framework obtained an accuracy of 91.02%, close to the accuracy level of state-of-the-art complex deep learning-based models. The results achieved show it is possible to develop a model more efficient and simple enough to be embedded in wearable devices. Additionally, the findings also show that HAR systems can be simplified by reducing the sampling rate and discarding irrelevant data. Finally, the conclusions highlight the key contributions, the research limitations, and provide recommendations for further research