Konark: A RFID based system for enhancing in-store shopping - PowerPoint PPT Presentation

Konark: A RFID based system for enhancing in-store shopping experience Swadhin Pradhan 1 , Eugene Chai 2 , Karthik Sundaresan 2 , Sampath Rangarajan 2 , and Lili Qiu 1 . 1 UT Austin 2 NEC Labs America WP WPA 2017, Mo MobiSys ‘1 ‘17 Ju June 19, 19, 2017 2017

Motivation • Immersive • More Insights • Personalized • Fine-grained .. • Faster … Retailer Consumer WP WPA 2017, Mo MobiSys ‘17 ‘17 June 19, Ju 19, 2017 2017 2

Overview • Creating a RFID based system to provide ▫ Better consumer shopping experience ▫ Richer retail analytics Queue-free checkout Items bought? Items interested ? Which items customers put in the cart The link ed at a particular time window ? Detecting which items users are browsing WP WPA 2017, Mo MobiSys ‘17 ‘17 June 19, Ju 19, 2017 2017 3

Our Work Building a RFID reader and antenna equipped shopping cart and developing algorithms to detect items inside the cart , and to detect users’ browsing interests on-the-go . WP WPA 2017, Mo MobiSys ‘17 ‘17 Ju June 19, 19, 2017 2017 4

Detecting Items Inside the Cart Reference Key idea Tags T h Algorithm sketch: Time Hi-dimensional Feature Space Tag IDs Yes K-Means Clustering Cart is Mobile ? Feature Values WP WPA 2017, Mo MobiSys ‘17 ‘17 Ju June 19, 19, 2017 2017 5

User’s Item Browsing Detection • Key idea : The phase variation of the tags of interacted items is higher. Algorithm sketch: Cart is Mobile ? Track phase variation No across tags Filter tags according to Find the tags with highest phase variation nearest reference tags WPA 2017, Mo WP MobiSys ‘17 ‘17 June 19, Ju 19, 2017 2017 6

Setup • Impinj Speedway Revolution RFID Reader ▫ Reads phase, RSSI, doppler of RFID tags ▫ 300 tags/second ▫ 50 channels, 902.75-927.75 (25 MHz bandwidth) • 6 dBi gain circular polarized Antennas • Dogbone Monza 6 RFID tags (Suited for 866-928 MHz reading) WP WPA 2017, Mo MobiSys ‘17 ‘17 June 19, Ju 19, 2017 2017 7

Experimental Setup Reference Tags Laptop The Cart Antennas RFID Reader WP WPA 2017, Mo MobiSys ‘17 ‘17 Ju June 19, 19, 2017 2017 8

Evaluation metrics § Modules : § Detecting items inside the cart. § Detecting user interest in a particular item. § Metrics : Ø Accuracy (What percentage of items predicted correctly ?) o False Positive % (What percentage of items are outside but tagged inside ?) o False Negative % (What percentage of items are inside but tagged outside ?) Ø Detection Latency (How much time it takes to detect ?) WP WPA 2017, Mo MobiSys ‘17 ‘17 June 19, Ju 19, 2017 2017 9

In-cart item Detection Accuracy Accuracy reaches ~100% after Detection latency of 60 seconds Accuracy remains good even after putting many items in the cart WP WPA 2017, Mo MobiSys ‘17 ‘17 Ju June 19, 19, 2017 2017 10

Detection Latency vs Accuracy False positive rate is not too high even in low detection latency and high number of items WP WPA 2017, Mo MobiSys ‘17 ‘17 Ju June 19, 19, 2017 2017 11

Item Browsing Detection 100 20s 30s 98 40s 96 For number of items ~ 200, 94 Accuracy to achieve ~100% accuracy , 92 we need only ~20 sec latency . 90 88 86 100 200 400 500 Number of items in the vicinity of the cart WPA 2017, Mo WP MobiSys ‘17 ‘17 June 19, Ju 19, 2017 2017 12

Key takeaways § Trade-off between detection latency and accuracy. § Speed and benefits compared to traditional self-checkout system or vision based system (Amazon Go). (NLOS/Occlusion). § I nfrastructure RFID solutions ( ShopMiner [SenSys ‘15] or CBID [Infocom ‘14]) which work with smaller number of tags , and lacks user level information . WP WPA 2017, Mo MobiSys ‘17 ‘17 June 19, Ju 19, 2017 2017 13

Future Works § Making the retail analytics richer . § Testing with multiple carts at different mobility. § Collaboration among shopping carts etc. § Field-testing the system in real shopping malls. WP WPA 2017, Mo MobiSys ‘17 ‘17 Ju June 19, 19, 2017 2017 14