Analysis of online and offline classification algorithms for human activity recognition using IMU sensors

Physical activity monitoring through machine learning, using data collected from wearable devices equipped with motion sensors and vital signs monitoring, such as heart rate, temperature, and blood oxygenation, has gained significant attention in sports and medical fields. This advancement enables real-time performance tracking and early detection of motor conditions. While offline classifiers achieve high accuracy, they cannot adapt to novel motion patterns; online (incremental) learners overcome this limitation. Although there are online learning algorithms, their application to Human Activity Recognition (HAR) remains limited. The challenge for offline and online approaches is generalizing activity detection based on time-series segments, relying solely on sensor data without additional information. This study analyzes the performance of offline and online algorithms for HAR by applying segmentation and feature extraction techniques and evaluating the adaptability of incremental learning models over time. The research follows a quantitative and prescriptive approach, employing various classifiers to address the problem of HAR. The results highlight the performance of the offline Autogluon Predictor applied to the PAMAP2 dataset, which the five best algorithms nearly tied and achieved average scores of 94% accuracy, 93% F1-score, 94% precision, and 94% recall, followed by CIF with lower results. For online learning the algorithms XGBoost Incremental performed best with 78% accuracy, 77% f1-score, 81% precision and 78% recall. While Adaptive Random Forest and BiLSTM also nearly tied in the online setting, achieving lower results compared to XGBI, this study shows promising results for both scenarios. It highlights that determining the most suitable learning paradigm (online or offline) is a decision for the data scientist, guided by the particularities of the problem and the necessary dynamism required for the scenario. By exploring different learning approaches for HAR and evaluating their effectiveness, this research contributes to the development of more adaptive and personalized systems with applications in health monitoring, sports, and medical diagnostics, fostering advancements in continuous and adaptive user data analysis.