Sensors-Based Human Activity Recognition Using Hybrid Features and Deep Capsule Network

Abstract

With the continuous advancements in artificial intelligence, human activity recognition (HAR) technologies have garnered widespread attention and found applications across diverse domains. Recently, various features and deep learning models are proposed for human activity recognition using sensors data. Though the existing models and features have achieved notable performance, their recognition accuracy needs to be enhanced and computational cost needs to be reduced. This research introduces a novel integration of various features and deep capsule network capable of real-time signal processing. To construct the HAR model, the information about patterns shape and symmetry of the sensor data distribution, and variations in frequency modulation was extracted. This information was given to the deep learning model (DeepCapsNet) that integrates several convolutional layers (CLs) and deep capsule network. The CLs in DeepCapsNet are used to process temporal sequences and deliver scalar outputs, while the capsule network is utilized to retrieve the equivariance which enhances the performance of HAR model. Lastly, the efficiency of the DeepCapsNet is comprehensively assessed in comparison to other baseline models using three benchmark HAR datasets. The average accuracy of DeepCapsNet for UCI HAR, WISDM, and PAMAP2 datasets are 97.6%, 98.5% and 99.9% respectively. This research findings revealed the effectiveness of the DeepCapsNet over the performance of the baseline models in terms of accuracy and computational cost. The feature selection and model optimization need to be further explored to enhance the performance of HAR model.