Human Actvity Recognition Based on Universal Time Series Features Using Smartphone Sensor Data

This study aims to effectively recognize human activity by utilizing sensor data built into smartphones.Human activity recognition (HAR), which automatically detects human behavior, can be categorized into two main approaches: vision-based and sensor-based methods.This study uses an approach that extracts handcrafted features from smartphone sensor data and recognizes them through machine learning.While most studies utilizing machine learning techniques concentrate on identifying a limited number of features suitable for HAR, this study proposes an approach that extracts universal features for activity recognition in both time and frequency domains, commonly used in signal processing.The classification is performed using Multi-Layer Perceptron (MLP) or Support Vector Machine (SVM).This method benefits from the adaptability of deep learning approaches, allowing the classifier to adjust based on the features without needing to change different features when the target activity class changes.Additionally, it is capable of efficient learning with a relatively small dataset, as it has fewer parameters to learn compared to deep learning methods.To validate the effectiveness of the proposed method, an experiment was conducted using the UniMib SHAR dataset, which is widely used in HAR research.The experimental results indicated that the proposed method outperformed existing studies in terms of recognition performance, both in the 5-fold cross-validation method and in the userindependent method.