Year: 2024
Type: dissertation
Abstract: The integration of artificial intelligence (AI) into daily life has brought significant advancements, especially in healthcare where Human Activity Recognition (HAR) systems play a vital role. These systems are designed to identify and monitor human activity in real-time, providing valuable insights into the physical behavior of subjects. With the development of AI, the field of HAR has also evolved from traditional methods to deep learning-based techniques for time series data. \nThis research applies various deep-learning methodologies by training and evaluating models utilizing various public benchmark datasets. By comparing the performances of various models, a comprehensive understanding of the efficiency and robustness of the approach is obtained. Additionally, the models are trained and evaluated on a newly constructed dataset collected with smartwatches, providing insights into the practicality and adaptability of each model in real-world scenarios. \nFurthermore, two ensemble architectures for multi-modal datasets are introduced, training and evaluating a dataset collected from both the accelerometer and heart rate monitor. Especially, an ensemble architecture, combining two DeepConvLSTM models, achieved a high F1-score. The findings demonstrate that the proposed model can extract precise and comprehensive information from multiple modalities by achieving superior performance compared to individual models trained with the dataset containing an accelerometer or heart rate monitor only.
Author: 김소영
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7027701638
