Effectiveness of ANN, LSTM, and Various Supervised Machine Learning Algorithms on Human Activity Recognition

Abstract

Plenty of healthcare applications (fitness tracking, sleep pattern monitoring, amount of calories burned, etc.), remote monitoring applications (patient, child or older people), security, and soft biometrics applications require automatic and precise activity recognition. So, this paper demonstrates the best approach among various approaches used for activity classification using the tri-axial gyroscope and accelerometer sensors, which are already integrated into smartphones. Some unique and distinguishable features are extracted using the acquired signals from these sensors. For activity recognition purpose, we have evaluated the performance of the Artificial Neural Network (ANN), Long short term memory (LSTM), and various supervised machine learning algorithms on a human activity recognition data set, UCI HAR dataset (obtained from the inertial sensors of a smartphone). Comparative results obtained on UCI HAR dataset shows that the maximum accuracy achieved so far is 98.57%. Therefore, we have analyzed the effectiveness of various methods using our feature extraction process and observed that 99.9% of accuracy can be achieved with the Ensemble classification method with bagging tree. © 2023 IEEE.