Human Activity Recognition Based on Machine Learning and Smartphone Sensors

Abstract

In recent years, the importance of Human Activity Recognition (HAR) technology has continued to grow. This study utilizes smartphone-embedded Inertial Measurement Units (IMUs) to collect real-world accelerometer and gyroscope data, constructing a multi-dimensional activity analysis framework. Through systematic evaluation of time-domain, frequency-domain, and time-frequency domain feature extraction methods, combined with SMOTE data balancing and PCA dimensionality reduction techniques, a hierarchical classification architecture was designed. This framework integrates machine learning models including logistic regression, Support Vector Machine (SVM), Random Forest, XGBoost, and KNN. The results demonstrate differentiated characteristics of classifiers in activity recognition tasks. By dynamically selecting classifier combinations adapted to different activity features, the system achieves flexible optimization of classification strategies. Validation reveals that the integration of time-frequency domain features with dimensionality reduction effectively enhances computational efficiency. The proposed hierarchical framework provides technical support for lightweight HAR systems on smartphone platforms, with its multi-feature domain fusion strategy being extendable to daily activity monitoring and health assessment scenarios. This research offers methodological references for human behavior understanding in wearable devices, emphasizing the critical role of feature engineering and classifier co-optimization in practical applications.