Implementation of Machine Learning Algorithms For Human Activity Recognition

Human Activity Recognition (HAR) is technically the problem of forecasting an individual's actions based on evidence of their gesture using sensors functioning as accelerometer and gyroscope. It plays a major role in contrasting sectors such as personal biometric signature, daily life monitoring, anti-terrorists along with anti-crime securities, medical-related applications, and so on. These days, smart phones are well-resourced with leading processors and built-in sensors. This comes up with the possibility to unfold a new arena of data mining. This paper signifies the analysis of HAR focused on data composed via accelerometer sensors of smart phones. Further, it illustrates the use of time-domain features which are acquired with the help of a windowing approach termed as overlapping. It is accompanied by a window size of 250ms along with overlapping of 25%. Numerous machine learning classifiers such as k-nearest neighbors, linear discriminant analysis, bagging classifier, gradient boosting classifier, decision tree, random forest, and support vector machine using three different kernels were practiced. The outcomes exhibit that random forest with 5-fold cross-validation imparts the highest accuracy (92.71%) in recognition of human activities.