Human Activity Recognition using EfficientNet for Wearable Sensor Data

Abstract

Human Activity Recognition (HAR) has attracted considerable attention from researchers due to its broad implications in healthcare, smart environments, and entertainment. HAR involves the usage of sensor devices and the Internet of Things (IoT). Human Activity Recognition (HAR) has emerged as a crucial application in health monitoring, necessitating the ongoing utilization of Smart phones, smart watches, and wearable devices to document and track patients' everyday activities. Deep learning (DL)-based algorithms have shown effective in predicting various human actions using time-series data collected from cell phones and wearable sensors. Deep learning-based methods have faced challenges when used to time-series data in activity recognition. The proposed methodology has the potential to effectively address those concerns. In this paper a novel approach is proposed using EfficientNet by adding various layers during the classification phase. Proposed method is evaluated by using UCI-HAR dataset and PAMAPS2 dataset, Class imbalance during training and testing phase is reduced by using various data augmentation methods. Proposed method achieved an accuracy of 99.98 on PAMAPS dataset thereby achieving significant results than state of art methods.