A Human Activity Recognition System Using Temporal Convolutional Network with Multi-Scale Temporal Attention Mechanism

In recent years, the common adoption of smart glasses has created new opportunities for human and computer interaction mainly in Human Activity Recognition (HAR) domain. The smart glasses can capture a collection of wearer's activities data with the help of built-in sensors and cameras. However, the development of HAR systems for smart glasses led to significant challenges such as the requirement of real-time processing and effective computation on resource-constrained devices. Therefore, this research aims to develop an embedded Artificial Intelligence (AI) powered HAR system for smart glasses using Temporal Convolutional Network (TCN) with Multi-Scale Temporal Attention Mechanism (MSTAM). Initially, the data is gathered from HAR dataset, a complete source of human activities. Then, the collected data is further preprocessed with the help of Dynamic Time Warping (DTW), as it aligns time-series signals of verifying lengths. After that, Haar Wavelet and Symlet Wavelet Coefficient Scattering (HW-SWCS) is employed to extract the features from preprocessed signals. Finally, TCN–MSTAM is introduced to recognize and classify the human activities. From the results, the proposed TCN-MSTAM model offered outstanding results in accuracy of 93.4% when compared with existing Sensors based on a Capsule Network (SensCapsNet).