Human Activity Recognition Using Smartphone Sensor Data: A Lightweight Benchmark Study

Abstract

Smartphone inertial sensors enable Human Activity Recognition (HAR) which serves as a fundamental task for mobile computing and digital health and context-aware systems because these devices offer both functional sensing abilities and extensive access to real-world environments. The study creates a journal-style lightweight benchmark for smartphone sensor-based HAR through a dataset which contains 4,999 samples and 561 numerical features and 15 subjects and six activity classes after data cleaning and restructuring. The proposed pipeline performs data repair before preprocessing and descriptive statistics and feature grouping and correlation analysis and principal component analysis (PCA) and ANOVA-based feature ranking and model benchmarking. The study tested four lightweight machine learning baselines which included Logistic Regression and Random Forest and k-Nearest Neighbors (KNN) and Gaussian Naive Bayes through an 80/20 train-test split that maintained class distribution. The Logistic Regression model delivered the highest performance with an accuracy of 0.973 and a macro F1-score of 0.9748 yet Random Forest and KNN models also achieved results above 0.95. The analysis of features demonstrated that gravity-based descriptors together with angle-based descriptors produced the most effective recognition results yet the error analysis showed that the most frequent classification errors occurred between sitting and standing positions. The research demonstrates that feature representations which have been designed with purpose maintain their competitive position when used in deployment-focused HAR systems.