Physical Human Activity Recognition Using Machine Learning Algorithms

With the rise in ubiquitous computing, the desire to make everyday lives smarter and easier with technology is on the increase. Human activity recognition (HAR) is the outcome of a similar motive. HAR enables a wide range of pervasive computing applications by recognizing the activity performed by a user. In order to contribute to the multi facet applications that HAR is capable to offer, predicting the right activity is of utmost importance. Simplest of the issues as the use of incorrect data manipulation or utilizing a wrong algorithm to perform prediction can hinder the performance of a HAR system. This study is designed to perform HAR by using two dimensionality reduction techniques followed by five different supervised machine learning algorithms as an aim to receive better predictive accuracy over the existing benchmark research. Correlation analysis (CA) and Principal component analysis (PCA) are used for feature reduction which resulted in 173 and 100 features respectively. Decision Tree, K Nearest Neighbor, Naive Bayes, Multinomial Logistic Regression and Artificial Neural Network algorithms were used to perform the classification task. The repeated random sub-sampling cross validation technique was used to perform the evaluation followed by a Wilcoxon signed rank test to evaluate the significance of the tests. The study resulted in ANN performing the best classification by achieving 97% of accuracy using the CA as feature reduction technique. The KNN and LR also provided satisfactory results and have received predictive results greater than the benchmark test. However, the decision tree and Naive bayes algorithms didn’t prove efficient