PTeR Model: A Computational Time Reduction and Performance Improvement Model for Recognizing the Associated Human Activity Using Smartphone Sensors

Human Activity Recognition (HAR) has been identified as a hotspot for finding the daily living activity of human beings. The researchers have proposed a substantial number of models for HAR. In addition, both time and performance domains greatly emphasize the experimental outcomes. The research challenges occurred whenever the HAR model produced the optimal results under these domains and found compelling temporal sensor activity data features. To address these research challenges, we have proposed a Performance and Time-efficient Recurrent (PTeR) model for learning and capturing the frequent occurrence of activity patterns. Further, the proposed model contains the recurrent networks for handling the temporal sensor data. The experimental findings regarding performance measurement and time reduction have been evaluated using the FLAAP dataset. We have used the F1 score (%) and precision (%) for the performance measurement. Moreover, total training time (sec.), saved training time (sec.), and their percentage employed for computing the time consumption. The impact of computational time reduction and performance improvement are evaluated. According to the experimental findings, the PTeR (RNN) has surpassed Random Forest (RF), CNN + RNN, CNN + LSTM, and has attained the F1 score of 95.04% and a precision of 89.95%. Further, the PTeR (GRU) saved 50.15% of the overall training time and used 221.38 s less than the CNN + RNN and CNN + LSTM models. Finally, the recognition rates of the PTeR model outperformed the comparative models developed using convolutional-aided and traditional learning algorithms.