Improvement of Human Activity Recognition (HAR) Performance by Utilizing LSTM in the Structure of Progressive Learning

Human Activity Recognition is a crucial area of research in many applications. However, training deep neural network-based models can be expensive and time-consuming. This paper introduces a progressive learning-based approach designed to not only reduce dimensionality but also maintain high accuracy. The proposed method replaces extensive sensor data with four extracted features, derived using a CNN-LSTM-based model trained on the WISDM dataset comprising six activity classes. The extracted features are fed into machine learning algorithms and achieve 97.73% accuracy.