A Hybrid Feature Selection Method for Human Activity Recognition

Abstract

Human Activity Recognition (HAR) is a highly active research field, as it is essential in various domains such as healthcare, surveillance, sports, robotics, etc. Many HAR systems are based on Machine Learning (ML) algorithms. However, deploying these systems is often a significant challenge due to their spatial and temporal complexity, especially in real-time applications where quasi-instantaneous recognition is imperative. This article proposes a novel hybrid feature selection approach that optimizes the execution time of a HAR system. The feature selection combines Cuckoo Search (CS) and Recursive Feature Elimination (RFE) algorithms. Subsequently, the selected features feed into various ML classification algorithms such as Support Vector Machine (SVM), Random Forest (RF), and Logistic Regression (LR) to recognize activities. The proposed approach significantly reduces the number of features and, consequently, the temporal complexity. The results of the experiment performed on the UCI-HAR dataset compete with those achieved in the state-of-the-art using fewer features.