Effect of Shallow and Ensemble Learning Models for Human Activity Recognition in Uncontrolled Environment

Human activity is one of the most exploited areas in the field of artificial intelligence and wearable sensor technology. Various users create different datasets in distinct environments to classify daily living activities using a suitable learning algorithm. Data collected in the simulated or controlled environment suffers in real-time activity recognition. We created a new HAR dataset in an uncontrolled environment and compared its result with different learning models. Various algorithms based on shallow and ensemble approaches are incorporated for the experimental study, and their effect is analyzed in detail. The importance of different data pre-processing steps are laid out with empirical evidence. We managed to get an average accuracy of 98.6% and 98.85% with random forest and eXtreme gradient boosting algorithms, respectively. Further, we concluded that the ensemble approach is more suitable for activity recognition and has a potential advantage over performance evaluation.