The ensemble methods play a vital role in machine learning for obtaining a high-performing model for the study dataset, and combining multiple classifiers to build a best-predictive model. On the other hand, Feature selection helps to remove irrelevant variables in the dataset in order to construct better predictive models. Therefore this research aimed to develop a robust model for activity recognition for multi-residents in smart homes using the ARAS dataset. The study employed Tree-based feature selection to cater to feature selection; two ensemble approaches, hard and soft voting, in line with five base learner classifiers: Logistic Regression (LR), Linear Discriminant Analysis (LDA), Naïve Bayes (NB), Random Forest (RF), and K-nearest neighbor (KNN), were applied to build the human activity recognition (HAR) model. The experimental results show that RF performed best compared to the rest of the classifiers, with an accuracy of 99.1%, and 99.2% in houses A and B, respectively. In comparison to prior findings, Feature Selection and ensemble methods enhanced prediction accuracy in the ARAS dataset.