Deep Learning Networks for Complex Activity Recognition Based on Wrist-Worn Sensor
Sakorn Mekruksavanich, Anuchit Jitpattanakul · 2023
Wearable smart devices, such as smartphones and smartwatches, offer great potential as platforms for automated human action identification. However, accurately monitoring complex human actions on these devices poses a challenge due to the presence of similarities in patterns across different actions. This occurs when distinct human actions exhibit comparable signal patterns or characteristics. The placement of motion sensors on the body plays a crucial role in detecting human behavior. Typically, wearable sensors placed at the trouser pocket or a similar location are used for this purpose. However, this positioning is not suitable for identifying actions involving manual gestures. To address this, wrist-worn motion sensors are employed to detect these specific behaviors. This study aims to investigate the effectiveness of deep learning models in accurately categorizing complex human actions using sensor data from wrist-worn devices. Nine deep learning models utilizing convolutional neural networks and recurrent neural networks were examined for their identification capabilities. The models were evaluated using the WHARF dataset, a publicly available benchmark dataset for human activity recognition. The investigation revealed that the proposed CNN-BiGRU model outperformed other deep learning models, achieving an accuracy rate of 87.20% and an Fl-score of 84.46%.