Real-Time Human Activity Recognition Using the H2O AutoML Framework and Imbalance Handling
Abstract:
Human Activity Recognition (HAR) is essential in ubiquitous computing scenarios such as healthcare monitoring, smart environments, and wearable technologies. This study introduces an optimized, real-time HAR framework utilizing the H2O AutoML platform, which automates model training, hyperparameter tuning, and performance benchmarking. We leverage the HARTH dataset, where high-frequency accelerometer data is preprocessed and segmented using sliding windows, followed by the extraction of temporal and statistical features. To address class imbalance in activity labels, we apply the Synthetic Minority Oversampling Technique (SMOTE), enhancing the robustness and generalization of the trained models. All modeling tasks are executed within the H2O AutoML environment, allowing efficient exploration of diverse algorithms and ensemble strategies. Realtime applicability is validated through latency analysis, confirming that selected models achieve inference times under 100 milliseconds. Model evaluation is based on accuracy, F1-score, logloss, AUC, and AUCPR, with top performers identified via the AutoML leaderboard. Results show that the proposed approach achieves state-of-the-art recognition performance while meeting real-time constraints, making it suitable for time-sensitive HAR applications.