Refined LSTM Network for Sensor-based Human Activity Recognition in Real World Scenario

Abstract

Sensor-based identification of human actions is an essential field of study in ubiquitous computing. This aims to facilitate the assessment or understanding of current occurrences and their context based on sensor signals. Activity recognition is employed in surveillance systems, patient health monitoring, and many other systems involving the interaction between human and intelligent wearable devices, including smartphones and smartwatches. The primary objective of this study work is to identify human behavior in the actual world. We proposed an improved long short-term memory network called RLSTM that uses a squeeze-and-excitation module to efficiently identify human actions and enhance action identification systems' interpretation. A publicly available real-world dataset known as REALWORLD16 was used to train and validate the model five times to analyze the proposed network. The proposed RLSTM achieved the highest accuracy of 98.04% and F1-score of 97.76%, as determined by several investigations.