A Video-Based Measurement Framework for Chewing-Event Detection Using 3D Facial Landmark Dynamics and sEMG-Based...

Area of Science:

• Biomedical Engineering
• Computer Vision
• Human-Computer Interaction

Background:

• Accurate measurement of chewing events is crucial for monitoring feeding behavior and masticatory function.
• Existing methods often require contact sensors, wearables, or manual annotation, limiting unobtrusive monitoring.
• There is a need for non-contact, automated methods for chewing event detection in naturalistic settings.

Purpose of the Study:

• To develop and validate a non-contact, video-based framework for detecting chewing events during natural eating.
• To utilize normalized 3D facial landmark dynamics and recurrent temporal modeling for chewing event detection.
• To assess the system's accuracy, robustness, and real-time applicability compared to existing methods.

Main Methods:

• A novel framework employing frontal facial video analysis was developed.
• Normalized 3D facial landmark dynamics were extracted and processed using recurrent neural networks.
• Physiologically grounded reference labels were derived from synchronized electromyography (EMG) during real-meal sessions.

Main Results:

• The video-based method achieved high accuracy on an independent test set, with a mean absolute error of 4.4 chews/session.
• The proposed approach demonstrated significantly lower counting errors compared to a rule-based video method, especially during concurrent activities.
• The system achieved real-time performance (mean latency < 9 ms on CPU, < 7 ms on CUDA) and was implemented as an Android application.

Conclusions:

• The non-contact, video-based framework offers an accurate and unobtrusive method for chewing event detection.
• The system demonstrates superior performance over rule-based video methods, particularly in complex eating scenarios.
• The real-time capabilities and mobile implementation support practical deployment for feeding behavior analysis.