Advancing Human Activity Recognition: A Novel WAECN-BO Approach for Distinguishing Highly Correlated Actions

Abstract

Human actions, when gauged through the lens of Human Activity Recognition (HAR), find numerous applications across healthcare, sports, and security sectors. Nonetheless, the intricacy of HAR becomes apparent when distinguishing akin actions poses a challenge. To tackle this issue, the present article introduces a pioneering method known as the Weighted Average Ensemble of Convolutional Neural Networks with Bayesian Optimization (WAECN-BO), which amalgamates five distinct Convolutional Neural Network (CNN) layer configurations. Notably, this method incorporates a fresh CNN layer designed to enable more intricate abstraction and optimizes its hyperparameters through Bayesian optimization. The evaluation of this method transpires on the UniMiB SHAR Database, a well-recognized benchmark dataset for HAR, focusing on actions with considerable resemblance. The findings reveal a remarkable accuracy rate of 99.98% across the entire dataset, surpassing established state-of-the-art approaches. Additionally, an analysis of the individual contributions made by each CNN layer configuration to the model's performance is conducted. This method, poised to enhance the accuracy of HAR systems across diverse domains, especially those dealing with actions that closely resemble each other, emerges as a promising advancement.

Internet of Medical Things and the Evolution of Healthcare 4.0: Exploring Recent Trends

“…Assisted living, an important form of long-term care, supports elderly as well as disabled individuals with daily activities, promoting comfort and minimizing hospitalization [72]. Telepresence and videoconferencing robotic solutions ensure continuous monitoring, fostering communication, and offering educational and therapeutic activities [73].…”