Physique-Driven Human Activity Recognition with Ensemble Learning

In this research paper, we address a common limitation in traditional Human Activity Recognition (HAR) methods, which often neglect the physical attributes of human subjects, such as height, weight. This oversight can lead to inconsistent performance across subjects with diverse physical characteristics. To remedy this issue, we introduce a novel Physique-based HAR approach designed to enhance accuracy and efficiency. Our method utilizes raw sensor data from built-in smartphone accelerometer and gyroscope sensors. After reprocessing the data, we create physique-based data sets by grouping subjects with similar physical attributes. We conduct a comprehensive analysis, comparing machine learning algorithm performance on both physique-based and conventional HAR datasets. Our findings not only identify the most suitable machine learning algorithms for HAR but also demonstrate the superior performance of the proposed physique- based method over traditional non physique approaches. We achieve an impressive 97.88% accuracy rate, validated on publicly available and custom-generated data sets. Additionally, we provide a detailed breakdown of activity-specific accuracy results, comparing them to recent benchmarks, underscoring the efficiency and effectiveness of our Physique-based HAR approach.