Uncertainty-Aware Human Activity Recognition: Investigating Sensor Impact in ML Models

Human activity recognition (HAR) using wearable sensors has gained significant attention in health monitoring, rehabilitation, and activity tracking. HAR enables the automatic identification of human movements based on sensor data, playing a crucial role in applications such as elderly care, fitness tracking, and medical diagnostics. The effectiveness of HAR systems relies on accurate data collection and robust classification algorithms to distinguish between different activities. However, the accuracy of machine learning (ML)-based HAR models depends on the quality of the ML algorithms and the reliability of sensor measurements. This study investigates the impact of accelerometer measurement uncertainty on the performance of HAR classification models. Using the HARTH dataset, which includes three-axis accelerometer data from two sensors placed on the thigh and lower back of several subjects, we introduce controlled perturbations to the raw data to simulate real-world measurement uncertainty and evaluate its effect on different machine learning models, including Random Forest (RF) and eXtreme Gradient Boosting (XGBoost). Our results highlight that the sensors' uncertainty and its distribution significantly affect the classification performance. The robustness of the models to different measurement noise is further analyzed, and the implications for designing HAR systems with unreliable or low-cost sensors are discussed. These findings provide valuable insights into the reliability of HAR systems in real-world conditions, guiding the development of more robust and uncertainty-aware ML-based classification models.