Deep Learning-Based Wearable Human Activity Recognition: Model and Performance Analysis
Xinyu Song, Hongyu Sun, Yanhua Dong, Ying Pei, Shuai Wang · 2024
Mobile wearable sensors, with advantages such as real-time monitoring, portability, data sharing and connectivity, increased user engagement, and multifunctionality, have gained significant attention from researchers. Their application in identifying, interpreting, and evaluating human behaviors is becoming increasingly prominent. However, inconsistencies in experimental conditions among researchers and the complex dependencies between components and modules within systems may impact the reproducibility, comparability, stability, and maintainability of studies. To address these challenges, this paper presents a unified solution for research on mobile wearable sensors in the context of human behavior. Specifically, it introduces a middleware model called HAR-IMB. This model encompasses input sensor data, feature transformation, model selection, and prediction results. To substantiate the model's efficacy, the study conducts performance analyses on different deep learning models using the WISDM dataset and UCI-HAR dataset as examples. The experimental results demonstrate promising outcomes across various models. Furthermore, the middleware model exhibits scalability. In the future, it can be further enhanced and refined to adapt to more complex experimental environments and application scenarios.