RFGo: A Seamless Self-checkout System for Apparel Stores Using RFID Carlos Bocanegra (Northeastern University) Mohammad A. (Amir) Khojastepour (NEC Laboratories America) Mustafa Y. Arslan (NEC Laboratories America) Eugene Chai (NEC Laboratories America) Sampath Rangarajan (NEC Laboratories America) Kaushik R. Chowdhury (Northeastern University)
AGENDA < 2 > 6 Conclusions Barcodes and 1 Alternatives RFID as RFGo 5 2 the key evaluation technology RFID-based 3 RFGo: vision, proposals 4 design and implementation
THE PAINFUL CHECKOUT PROCESS < 3 > Only 23% of consumers are satisfied with the checkout process 1 13% sees the checkout burden as the decisive factor to switch to another store 1 [1] Forrester. 2018. Consumers Cringe At Slow Checkout. Forrester Opportunity Snapshot: Digimarc August 2018 (8 2018). https://www.digimarc.com/resources/forrester-study
ALTERNATIVES TO BARCODE < 4 > 1. Computer vision & Deep Learning Amazon Go : Cameras and sensors (sensor fusion) Extensive hardware resources Dense camera deployment Privacy concerns 2. RFID based checkout [1] Clresearch. 2018. ” Amazon Go and the Emergence of Sentient Buildings: How It Works and What Its Impact Will Be ,” April 2018 (4 2018). http://www.clresearch.com/research/detail.cfm?guid=6A608036-3048-78A9-2FB3-4E6295D65919
RFID, WHY? < 5 > It is possible to build a seamless checkout system based on RFID ● it requires all items to be tagged 1. Use cases of RFID in retail sector is on the rise ● Inventory management, Reduce out-of-stock items, Tracking items at the warehouses 2. Cost per tag is low 3. Governments embrace this technology , i.e. Japan 2025 initiative 4. RFID is already in place for some major apparel retailers
STATE-OF-THE-ART RFID CHECKOUT SYSTEMS < 6 > Cage-based Slot-based Bin-based Surface-based Handheld Effortless Large area Unbarricaded Fast
OUR VISION FOR A SELF-CHECKOUT SYSTEM < 7 > EA Top CA No Unbarricaded manual checkout view effort area Checkout CA area Large High WA Wait area checkout Speed area checkout Customers and RFID items Exit EA WA area
BACKGROUND ON RFID < 8 > RFID basic components CHIP Gen2 configurations RFID And READER POWER TAG MEMORY Center Frequency : 900 MHz QUERY Bandwidth : 1 MHz approximately DATA ID Reader-To-Tag encoding : PIE ID Backscattering Tag-To-Reader encoding : FM0 ANTENNA R<->T communication phases FM0 modulation
CHALLENGES IN RFID < 9 > Blind Spots Collision Position Uncertainty ❏ Illumination ❏ Mobility 6 1 N RN16 R P O W E R RN16 ❏ Orientation ❏ Non-stationary environment ❏ Coupling
RFGo - OUR PROPOSED SYSTEM < 10 > Self-checkout vision 1. Physical structure 2. Custom-built multi-antenna reader 3. Tag classification via supervised learning
RFGo - 1. PHYSICAL STRUCTURE < 11 > Unbarricaded and large CA with no need for manipulating the items IR sensors to assess occupancy within the CA Session : chain of operations including entry, scanning, classification and output 10 antennas, 6 covering the CA and 4 covering the outer region
RFGo - 2. CUSTOM-BUILT READER < 12 > Collisions Low Reading Rate Slow checkout ● Conventional methods to resolve collisions, effective but not in real-time ● RFGo - Exploits diversity in reception ● Multi-antenna commercial readers - TDMA-TX/RX do not exploit diversity
RFGo - 2. SELECTING THE RN16 TO ACK < 13 > Packet Delivery Ratio (PDR) ● RN16 lack error detection mechanism, e.g., CRC ● SINR is a post-decoding metric and is not available before decoding ● Can we find a pre-decoding metric which follows the idea of SINR? Our solution: Interference Metric (IM)
RFGo - 2. INTERFERENCE METRIC (IM) < 14 > ● Revisiting differential decoding FM0 symbols 20dB SNR 1 Tag 1 cluster 2 clusters (bits 1/0)
RFGo - 2. INTERFERENCE METRIC (IM) < 15 > 10dB SNR 30dB SNR / 5dB SIR 30dB SNR / 5dB SIR 1 Tag 2 Tags 3 Tags Std/mean Std/mean Std/mean = 0.28 = 0.53 = 0.33
RFGo - 2. INTERFERENCE METRIC (IM) < 16 > *IM -> IM Policy ( IMP ) Packet Delivery Ratio (PDR)* IM assesses the RN16 during decoding IM does not incur in extra computation cost IM is easily parallelizable across the RX-chains SINR (Post decoding metric) IM (Pre decoding metric)
RFGo - 2. CUSTOM RFID READER IMPLEMENTATION < 17 > UBX daughterboard Octoclock for for single frequency And TX-chain time sync Raspberry Pi USRP X310 with controls the TwinRX active TX daughterboards antenna for RX-chains through MUX Smartrac Battery-less UHD RFID tags [1] Nikos Kargas, Fanis Mavromatis, and Aggelos Bletsas. 2015. “ Fully-Coherent reader with commodity SDR for Gen2 FM0 and computational RFID, ” IEEE Wireless Communications Letters 4, 6 (2015), 617–620. https://doi.org/10.1109/LWC.2015.2475749
RFGo - 3. TAG CLASSIFIER < 18 > Neural Network formed by 3 hidden intermediate layers #readings_RX 1 RSSI_RX 1 Inside CA #readings_RX 2 RSSI_RX 2 ... ... ... ... Outside CA #readings_RX N Training stage uses a RSSI_RX N wide range of orientations RSSI and # of readings as soft and locations in the 3D features for classification plane
RESULTS - BENEFITS OF IM < 19 > IM impact on PDR Packet Delivery Ratio (PDR) The fraction of slots that results in a correctly decoded EPC over the total number of slots. ● 6RX and 1TX. No blind spots. ● Slotted aloha saturates at 38%. ● FP reaches 52% resolving collisions ● Using the majority via MVP : 62% ● IMP wisely selects RN16 and reaches 77% PDR.
RESULTS - RECEIVER DIMENSION < 20 > IM impact on PDR and MDE Packet Delivery Ratio (PDR) The fraction of slots that results in a correctly decoded EPC over the total number of slots. ● IMP with 1TX and variable number of RX antenna ● PDR increase from 50% to 73% with 6 antennas. It saturates after.
RESULTS - BENEFITS OF DISCOVERY RATE < 21 > IM impact on Discovery Rate Discovery rate The percentage of the unique EPCs that have been decoded per unit of time. Variable TX, 1 RX Variable TX, 6 RX using IMP ● Multi-TX helps ● 2 RX helps speeding discovery by dealing with Blind almost 2x. spots but is slow . ● 6RX achieve full discovery under 1 second .
RESULTS - DEFINING THE CA AND GUARD AREA < 22 > Unified Cube RFGo inside/outside Inside-only The features: spillover readings go features of 6 inches 54’’ far from considerable shrinkage of the the CA. A spillover classifier is needed Untrained Inside-only features Inside-only features
RESULTS - PRECISSION AND RECALL < 23 > Deployment scenario High precision when RFGo does not include an outside tag in the customer cart. High recall when RFGo detects all the items that is in the customer cart Experiment with multiple orientations 872 tags Recall of 99.79% Precision of 99.77% Horizontal Vertical Random Experiment with Volunteers Recall of 99.68% Precision of 99.81%
CONCLUSIONS < 24 > 1. RFGo, a first-of-its-kind self-checkout system based on RFID 2. RFGo enhances customer experience with its effortless, open and unrestricted design 3. The multi-antenna framework increases the reading rate from 50% to 77% 4. The Supervised learning classifier achieves 99.79% precision and 99.77% recall
< 25 > THANKS
Recommend
More recommend
