Stacked LSTM Deep Neural Networks for Accurate Recognition of Prayer Activities with Smartphone Sensors

Smartphones have become an innovative technology with their ever-increasing processing, sensing, and networking capabilities, and they play a significant role in human activity recognition. Sensors embedded in the smartphones enable the recognition of human activities using machine and deep learning techniques. This study focuses on a specific type of human activity recognition that explicitly monitors and recognizes daily Muslim prayers. A novel deep learning framework is proposed to accurately recognize prayer activities: Stacked long short-term memory (LSTM) and bidirectional LSTM (BiLSTM) deep neural networks are utilized to recognize user motion through smartphone sensors. Stacked LSTM and BiLSTM deep learning techniques solve the vanishing gradient problem by preserving the relevant prayer activities performed by the user at each time step. Experiments were performed on two datasets collected using accelerometer and gyroscope sensors for the dawn prayer (13 activities) and the noon prayer (26 activities). The experimental results show that three-layer stacked LSTM and two-layer stacked BiLSTM deep learning techniques outperform state-of-the-art approaches in recognizing all prayer activities with an overall accuracy of 100% for dawn prayer data and 99% for noon prayer data. The hierarchy of hidden layers enables more precise recognition of complex activities. Another unique feature of the proposed system is that it accurately recognizes the missing activities during prayer. The proposed framework will assist both healthy people and people suffering from dementia or Alzheimer's disease who forget or repeat their activities due to memory loss.