Improving Inertial Sensor-based Human Activity Recognition using Ensemble Deep Learning

Sensor-based human activity recognition (S-HAR) is a famous study focusing on detecting human physiological actions by interpreting various sensors, especially one-dimensional time series information. Typically, S-HAR machine learning methods were developed using handcrafted characteristics. Unfortunately, this is a complicated process that involves feature engineering and a high level of domain knowledge. Due to the development of deep neural networks, classification techniques could efficiently handle relevant characteristics from raw sensor data, resulting in enhanced classification outcomes. In this study, we describe a unique method for S-HAR based on ensemble deep learning with sensor nodes connected to the waist, chest, leg, and arm. Implementing and training three deep learning networks is performed using a publically available dataset, including wearable sensors from eight human actions. The findings demonstrate that the proposed Ens-ResNeXt model provides the maximum accuracy and F1-score, which is superior to existing techniques.