When2Trigger: Evaluation Trade-offs in Vision-based Real-Time Eating Detection Systems.

Area of Science:

• Human-Computer Interaction
• Biomedical Engineering
• Behavioral Science

Background:

• Wearable sensors (cameras, thermal) aid real-time eating detection for meal logging.
• False positives from confounding gestures (e.g., hand movements) reduce accuracy.
• Balancing detection accuracy and delay is crucial for effective user feedback systems.

Purpose of the Study:

• To develop a real-time, hand-object-based method for automated eating and drinking gesture detection.
• To identify the minimum number of gestures for reliable eating episode detection.
• To reduce false positives in eating detection using low-power thermal sensing.

Main Methods:

• A novel method combining hand motion, object-in-hand recognition, and thermal sensing.
• Evaluation on 36 participants, with 28 wearing a camera in free-living conditions for up to 14 days.
• Analysis of gesture count and time-to-detection for eating episodes.

Main Results:

• Accurate eating episode detection achieved using 10 gestures or within 1.5 minutes.
• An F1-score of 89.0% demonstrates high detection performance.
• The method effectively reduces false positives compared to prior approaches.

Conclusions:

• A reliable, real-time eating detection system is feasible using wearable sensors.
• The developed method offers a balance between detection accuracy and timely feedback.
• Findings provide guidelines for designing interventions for problematic eating behaviors.