Hybrid Deep Learning-Based Human Activity Recognition (HAR) Using Wearable Sensors: An Edge Computing Approach

Due to the growth of Internet of Things (IoT) and advanced sensing based technologies have enabled the development of the miniature-based system. In recent years, the use of wearable and mobile sensors for Human Walking Gesture Recognition has become more popular in various applications, including health care, surveillance, robotics, and industry. The recent growth of edge computing technology for industry 4.0 has provided the opportunity to design the low power and less computationally expensive devices. The edge computing devices cannot support heavy computation and provide great efficiency by reducing the network size and communication latency. Deep learning algorithms have recently demonstrated high performance in HAR. However, the deep learning (DL) models require very high computation systems, which make them ineffective when used on edge devices. In this research, a hybrid deep learning-based model is trained to recognize the various gestures. Three deep learning-based models, namely one-Dimensional Convolutional Neural Network (1D-CNN), Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM), and CNN-Gated Recurrent Unit (CNN-GRU), are designed to test the various human mobility gestures. The WISDM, PAMAP2, and UCI-HAR benchmark datasets were used to assess these models. Among the three datasets, the best accuracies of the models are 99.89%, 97.28%, and 96.78%, respectively, achieved for CNN-LSTM hybrid model. In future, the work can be extended to design an end-to-end edge computing application using Arduino Nano 33 BLE Sensing microcontroller board. The compressed deep learning model will be fused on the Arduino Nano board to recognize various human motion gestures. The research demonstrates the classification of various HAR gestures using hybrid deep learning models.