Eat-Radar: Contactless Dietary Monitoring with FMCW Radar
Imagine a dinner table where the most attentive observer isn’t a person, but a small sensor tucked into the corner of the room. It doesn't see your face or record your voice; instead, it "feels" the ripples of your movement through high-frequency radio waves.
For millions living with chronic metabolic or cardiovascular diseases, managing health starts with a tedious log of what and how they eat. Traditional tracking is intrusive, relying on cameras that compromise privacy or wearable sensors that require constant charging.
A new study proposes a "contactless" future for dietary monitoring using Frequency Modulated Continuous Wave (FMCW) radar—a technology that can count every bite you take without ever needing a lens.
The Eat-Radar Study
Researchers utilized a 76 GHz radar to monitor N = 70 participants over 1,155 minutes of mealtime.
Unlike previous systems that merely guess meal times, this AI-driven approach, dubbed Eat-Radar, can distinguish the micro-movements of utensils like forks, spoons, or chopsticks.
By using a specialized 3D Temporal Convolutional Network with Attention (3D-TCN-Att), the system achieved high accuracy:
- Segmental F1-score of 0.896 for eating gestures
- Segmental F1-score of 0.868 for drinking gestures
The 3D Range-Doppler Cube Breakthrough
The key innovation is the radar's ability to preserve 3D Range-Doppler Cubes.
While 2D maps often lose spatial relationships, this 3D architecture captures the exact velocity and distance of gestures. This enables high accuracy even during confounding actions like checking phones or touching faces.
Accuracy varied by utensil:
- Chopsticks: F1-score of 0.920 (most precise)
- Hand-feeding: F1-score of 0.813 (slightly more elusive)
Benefits of Contactless Monitoring
For the average person, this technology means:
- Passive data collection: Health data is gathered invisibly without user intervention.
- No wearable devices: Eliminates the need for watches or sensors that require charging.
- Enhanced privacy: No video footage, avoiding privacy concerns associated with cameras.
- Precise identification: Accurately detects eating and drinking events, filtering out stationary clutter.
Current Limitations
The technology is not yet perfect:
- False positives: Deceptive movements, like wiping the mouth with a napkin, can occasionally trigger incorrect detections.
- Single-person optimization: The current model is designed for one person; it struggles in crowded settings like restaurants.
- Computational latency: The advanced 3D model has a latency of 1.062 seconds for processing 40 seconds of data—slower but still acceptable for real-time monitoring.
As the research team advances toward multi-person scenarios, this "invisible nutritionist" could become integral to modern preventative medicine.
This summary is based on: "Eat-Radar: Continuous Fine-Grained Intake Gesture Detection Using FMCW Radar and 3D Temporal Convolutional Network with Attention," published in IEEE Journal of Biomedical and Health Informatics (2023).