Smartwatch-based Eating Detection and Cutlery Classification using a Deep Residual Network with Squeeze-and-Excitation Module

• Published in: 2022 45th International Conference on Telecommunications and Signal Processing (TSP) pp. 301 - 304
• Main Authors: Mekruksavanich, Sakorn, Jantawong, Ponnipa, Jitpattanakul, Anuchit
• Format: Conference Proceeding
• Language: English
• Published: IEEE 13.07.2022
• Subjects: Behavioral sciences; cutlery classification; Deep learning; deep residual network; eating detection; Machine learning algorithms; Measurement; Signal processing; Signal processing algorithms; smartwatch sensors; Telecommunications
• DOI: 10.1109/TSP55681.2022.9851333

Several machine learning and deep learning algorithms have been developed to tackle the human behavior detection issue, emphasizing everyday tasks. Nevertheless, an intriguing and challenging human activity recognition (HAR) subject involves more complicated human behaviors, including eating-related actions. This study provided a smartwatch-based eating identification system based on hand-movement data. In the framework, we introduced a deep residual network named the ResNet-SE model that enhanced detection capability by using the benefits of shortcut connections and squeeze-and-excitation units. In addition, the effectiveness of standard deep learning models (CNN and LSTM) is evaluated and compared to that of the proposed model. In addition, we investigate the use of wristwatch sensor data for categorizing six kinds of food cutlery. To validate the correctness of the model, the accuracy, F1-score, and confusion matrices of the HAR metrics are applied to the EatingDetectionIJS dataset for evaluating the proposed frame-work. Experimentation findings demonstrate that the ResNet-SE model surpasses existing deep residual models, achieving the greatest F1-score of 91.81% for eating identification and 91.43% for cutlery categorization.