Achieving High-Accuracy Human Activity Recognition Using BERT-Based Classwise Ensemble Models
K. M Nafiur Rahman Fuad, Mumtahina Ahmed, Md. Golam Rabbani Abir, Md Anisur Rahman, Md Zakir Hossain Zamil, M. F. Mridha, Jungpil Shin · 2024
Human Activity Recognition (HAR) plays a vital role in ubiquitous computing and has a wide range of applications in healthcare, sports, and human-computer interaction. The study presents an innovative way of leveraging the BERT (Bidirectional Encoder Representations from Transformers) architecture to enhance the accuracy of HAR systems. By utilizing a classwise ensemble method, we effectively harness the capabilities of BERT for sequence classification, adapting it to the multi-dimensional time-series data characteristic of HAR tasks. Our model was trained and evaluated on the UCI HAR dataset, which consists of data collected with sensors from accelerometers and gyroscopes which recorded activities like walking, walking upstairs, sitting, walking downstairs, sitting, standing, and lying down from 30 participants. Our approach achieved outstanding performance through extensive experimentation, attaining 98% accuracy, 97% precision, 98% recall, and 97% F1-score. The results portray the robustness of the BERT-based classwise ensemble method in capturing complex patterns of human activities, outperforming traditional HAR models. This progress opens the door to more precise and dependable HAR systems, carrying substantial potential to improve user experiences across diverse application areas.