Ensemble of Multiple Classifiers for Accelerometer Based Human Fall Detection

Fall is a major health threat for the elderly. Automatic detection of human fall using machine learning algorithms is necessary to reduce the consequence of human fall. In the proposed method, nine different features resultant, variance, standard deviation, root mean square, Euclidean norm, skewness, kurtosis, and geometric mean have been calculated from accelerometer data. Stacking based ensemble of Naive Bayes, KNN, J48, and Random forest classifier has been used to discriminate fall activities for ADL. The performance of each classifier has also been calculated for comparing the performance of each classifier with a stacking-based ensemble method. It was found that even some individual classifier gives better result in sensitivity or specificity but stacking based ensemble of these classifier gives better results in a combination of sensitivity and specificity. With stacking-based ensemble classifiers, 89% sensitivity and 95% specificity have been achieved which is higher than the performance of individual classifiers.