i-Care Project Dietary-Aware Dining Table: Observing Dietary Behavior over Tabletop Surface Keng-hao Chang, Shih-yen Liu,Toung Lin, Hao Chu, Jane Hsu, Polly Huang, (Cheryl Chen) i-space Laboratory National Taiwan University 1 What is it? • A dietary-tracker built into an everyday dining table – Track what & how much you eat over tabletop surface • Motivation – We are what we eat – Food choices affect long-term & short-term health • Show a demo video 2 Smart Everyday Object • Digital-enhanced everyday objects – Provide digital services • Support natural human interactions – Natural human interactions = inputs to digital services • Goals – Providing digital services without (users) operating digital devices → better usability – Human-centric computing: technology adapting to users rather than users adapting & learning about technology 3 Outline for Reminder of Talk • • • • • • Related work Approach Assumptions & Limitations Design & Implementation Experimental Evaluation Future work 4 RelatedWork • Dietary trackers – Shopping receipt scanner (GaTech) – Chewing Sound (ETH) – My food phone (startup) • Intelligent surfaces – Load sensing table (Lancester) – Smart floor (GaTech, NTU) – Posture Chair (MIT) • What’s new here? – Accuracy – Fine-grained tracking – Simultaneous concurrent interactions 5 Contribution claims • It is a fine-granularity (automated) dietary tracker. – It can track multiple concurrent interactions from multiple individuals over the same tabletop surface. • People usually don’t eat alone • It is an enhanced loading sensing table. 6 General Approach • RFID tags on food containers • Two sensor surfaces on table – Each surface is made of cells – RFID reader surface • Detect RFID(s) in each cell – Weighting surface (load cells) • Measure weight change in each cell • Track the food path from container(s) → container(s) → mouth using these two sensor surfaces 7 Assumptions (Limitations) • Closed system rather than open system. – Food transfers among tabletop objects and mouths, no external objects and food sources • • • • • Users identified by personal containers (personal plates and cups) Food containers tagged with RFID tags No cross-cell objects No leaning their hands on the table Not a mobile tracker 8 Single Interaction Example • Bob pours tea from the tea pot to his personal cup, and drinks it • Detect tea transfer from one container to another container 1) Identify the presence & absence of containers • RFID tags on containers • tag-food mapping 2) Track tea transfer • Weight change detection • Weight matching algorithm 9 Single Interaction Example • Bob pours tea from the tea pot to personal cup, and drinks Put on tea pot. it •RFID tag appears •Weight increases ∆w3 Pour tea! • |∆w3 - ∆w1 | ≈ ∆w2 Pick up tea pot. Pour tea? •Weight increases ∆w2. • RFID tag disappears •Weight decreases ∆w1 10 Single Interaction Example • Bob pours tea from the tea pot to personal cup, and drinks it Put on cup. •RFID tag appears. •Weight increases ∆w2. Drink tea? (only if no match) • Amount | ∆w2 - ∆w1 | Pick up cup. • RFID tag disappears. •Weight decreases ∆w1. 11 Concurrent Interactions Example • Bob pours tea & Alice cuts cake Cut cake •Weight decreases ∆w2 Pour tea? Cut cake? •Weight change ∆w Pour tea •Weight increases ∆ w1 12 Concurrent Interactions Example • Multiple, concurrent person-object interactions – The larger the cell, the higher the possibility of concurrent interactions over a cell – Cell size = average size of container – Reduce the possibility of concurrent interactions over one cell 13 Design Architecture Applications (Dietary-aware Dining Table) Dietary Behaviors Behavior Inference Engine Intermediate Events Event Interpreter Sensor Events Weight Change Detector Object Presence Detector Weighing surface (weighing sensors) RFID Surface (readers) Tag-object mappings Common sense semantics 14 Inference Rule Dietary behaviors Behavior Inference Rules Transfer(u, w, type) Weight-Change(Object-i1, Δw1) ∩ (Δw1< 0) ∩ Weight- Change (Object-i2, Δw2) ∩ (Δw2 > 0) ∩ Contains(Object-i1, type) ∩ Owner(Object-i2, u) ∩(|Δw1 +Δ w2 |< ε) →Transfer (u, Δw2, type) Eat(u, w, type) Weight-Change(Object-i, Δw) ∩ (Δw<0) ∩ Contains(Object-i, type) ∩ Owner(Object-i, u)→ Eat(u, -Δw, type) 15 Experimental setup • 2 Dining settings – Afternoon tea – Chinese-style dinner • 2 Parameters – # of participants – Predefined vs. Random Keng-hao Sequence A Willy 16 Experimental Results Scenarios Dining # Activity Scenarios Users Sequence #1 Afternoon tea #2 Afternoon tea #3 Afternoon tea #4 Chinesestyle dinner 1 Event Statistics Time # of Average Duration Dietary Activity (seconds) Behavior Interval Predefined 73 12 6.08 Results Behavior Recognition Accuracy 2 Predefined 162 24 6.75 100% 2 Random 913 78 11.71 79.49% 3 Random 1811 162 11.18 85.8% 100% 17 Predefined Activity Sequence Afternoon Tea (Single User) 1. 2. 3. 4. 5. 6. cut a piece of cake and transfer it to the personal plate; pour tea from the tea pot to the personal cup; add milk to the personal cup from the creamer; eat the piece of cake from the personal plate; drink tea from the personal cup; add sugar to the personal cup from the sugar jar. Afternoon Tea (Multi-users) 1. A cuts cake and transfers it to A’s personal plate; 2. B pours tea from the tea pot to B’s personal cup; 3. A pours tea to A’s personal cup while B cuts a piece of cake and transfers it to B’s personal plate; 4. A adds sugar from the sugar jar to A’s personal cup while B adds milk from the creamer to B’s personal up; 5. A eats cake and B drinks tea; 6. B eats cake from B’s personal plate while A drinks tea from A’s personal cup; 7. A pours tea from the tea pot to both A’s and B’s personal cups. 18 Activity Recognition Accuracy in Scenario #3 Dietary Behavior Transfer event Eat event # of Actual Events 41 37 Recognition Accuracy 70.73% 89.19% 19 Causes of Misses in Scenario #3 Causes of misses # of misses of transfer events 6 # of misses of eat events 2 Total Weight matching threshold Slow RFID sample rate Touching table 2 3 1 0 0 2 2 3 3 Total of misses 12 4 16 Event interference within the weighing cell’s weight stabilization time 8 20 Activity Recognition Accuracy in Scenario #4 Dietary Behavior Transfer dish A events Transfer dish B events Transfer dish C events Transfer rice events Transfer soup events Eat events # of times 19 29 23 12 19 60 Recognition Accuracy Weight Accuracy 73.68% 79.31% 82.61% 68.42% 78.75% 79.19% 83.33% 84.21% 88.33% 81.88% 80.16% 91.23% 21 Causes of Misses in Scenario #4 Causes of misses Segmented weight-change events Eating before transferring food to personal containers Weight matching ambiguity Touching table Slow RFID sample rate Total of misses # of misses of transfer events 5 5 # of misses of eat events 0 5 Total 7 3 3 0 2 0 7 5 3 23 7 30 5 10 22 Conclusion • It is a smart object and a smart surface • It supports natural user interface • It supports fine-grained dietary tracking at individual level • It is about human-centric computing • Accuracy can be improved further 23 FutureWork • Improving recognition accuracy • Removing constraints (assumptions) • Persuasive computing – Encourage balanced diet – Encourage proper amount of diet 24 Questions & Answers ThankYou 25