Slide 1

Frogeye: Perception of the Slightest Tag Motion Lei Yang, Yong Qi, Jianbing Fang, Xuan Ding, Tianci Liu, Mo Li Tsinghua University, Xian Jiaotong University young@tagsys.org 2014-5-2 INFOCOM

Slide 2

Background RFID technology www.barcoding.com www.datasoft.se www.eff.org www.kennedygrp.com www.forrester.com TAGSREADERApplications

Slide 3

RFID Overview Portal Conveyor/Assembly line Access control Livestock Payment devices Logistics Passport Automobile immobilizers 55

Slide 4

MOTIVATION SECURING VALUABLE OBJECTS The most common solution is to equip artifacts with various security sensors, such as displacement sensor, tension sensor, vibration sensors and so on. As long as the artifacts are moved, alert is reported. These sensors are very expensive and difficult to be deployed. Camera surveillance is another attractive option Suffer from dead corners and dependence of the light. The most common solution is to equip artifacts with various security sensors, such as displacement sensor, tension sensor, vibration sensors and so on. As long as the artifacts are moved, alert is reported. These sensors are very expensive and difficult to be deployed. Camera surveillance is another attractive option Suffer from dead corners and dependence of the light. The Art of Securing Pricelessness

Slide 5

MOTIVATION MINING CONSUMERS BEHAVIOR What are the really popular products? In an effort to help supermarkets understand their consumers shopping behaviors, a large number of data mining techniques have been studied. However, those technique are confined to the purchased data. In most of time, the consumer takes their interested goods off the shelf for details but does not purchase them finally due to price. RFID technology offers an opportunity to collect these behaviors.

Slide 6

How to perceive the tags motion? At first glance, there is no any connection between the above two scenarios. Actually, both of them focus on the surveillance of tag motions: The first needs an alert when valuable objects are moved; the second requires behavior records when the products are taken off the shelf. Our goal is to perceive the tags motion to determine whether the object is moved. Our approach is not for localization

Slide 7

Opportunity Being hypersensitive

Slide 8

Challenge The Weak Stability Observation 2: The result is not as stable as expected, because the value occupies several units even when the tag remains in a same distance. We call this phenomenon weak stability. Observation 2: The result is not as stable as expected, because the value occupies several units even when the tag remains in a same distance. We call this phenomenon weak stability.

Slide 9

Which causes the weak stability Which causes the weak stability ? Thermal vibration: The electronic components thermal noise brings strength changes. Interference: when the strength is interfered, its changes are as significant as when the tag is moved. It is easy to mistakenly consider a stationary object moved.

Slide 10

Modeling the Thermal Vibration Gaussian Model : We believe this model is reasonable because a lot of natured phenomena follows the Gaussian distribution, especially thermal noise from internal electronic components, which mainly contribute the vibration.

Slide 11

Modeling the Interference This phenomenon is mainly explained by the multipath effect. There exist several paths for the backscattered signal propagating from tag to reader. The signal strength propagating through different paths varies a lot due to the path length. When the interference object gets close to the tag, it may block some propagation paths and leads to the propagation jumping among the multiple paths, resulting in the strength transmission from one level to another.

Slide 12

Modeling the Interference From a long-term perspective, the strength exhibits multimodal characteristics where the distribution is likely composed of multiple Gaussian models.

Slide 13

Basic Idea Our basic idea is to detect the significant changes of the backscattered signal for perception of tag motion. There is a high probability that the tag moved when its strength changed significantly. We find our problem is very similar to the foreground detection in computer vision, which is to segment the foreground pixels that significantly differ between the last image of sequence and the previous images.

Slide 14

Workflow

Slide 15

Preprocessing

Slide 16

Strength Image Construction In the image, each row is uniquely mapped to a same tag. The mapping fashion between the tags and rows is arbitrary as long as their mapping remains constant during the processing. Each column represents a read cycle. The whole image contains a total of m columns. Formally, given a strength image, the element x_ij represents a read strength from the tag i collected in the j^th read cycle of the frame. In the image, each row is uniquely mapped to a same tag. The mapping fashion between the tags and rows is arbitrary as long as their mapping remains constant during the processing. Each column represents a read cycle. The whole image contains a total of m columns. Formally, given a strength image, the element x_ij represents a read strength from the tag i collected in the j^th read cycle of the frame.

Slide 17

Why we convert the strength flow to a visual image? No any connections between them??

Slide 18

RATIONALE BEHIND Optical System RFID System

Slide 19

MOTION DETECTION IN COMPUTER VISION Frame Differencing The result is interesting and inspiring

Slide 20

MoG based Foreground Detection Background learning Background detecting frames The details can be referred to the paper.

Slide 21

Motion Refining collateral motion

Slide 22

Implementation & Evaluation We deploy one reader and 100 tags our noisy office room to evaluate the false positives. We attach tags on a toy train to measure the false negatives. The train moves along an oval track in a constant speed.

Slide 23

Evaluation the accuracy is up to 92.34% while the false positive is suppressed under 0.5%.

Slide 24

Sensitivity The average Minimum Perception displacement equals 6cm.

Slide 25

Evaluation

Slide 26

Slide 27

Conclusion In this paper, our major contributions are summarized as follows: We conduct statistical analysis of strength collected in a real-life office, showing that the strength are hypersensitive to tags positions, but suffers from weak stability where the strength values are highly clustered in a small range due to thermal noise, and enhanced or weakened due to multi-path effect. We present a MOG method to characterize the weak stability. We propose Frogeye, to perceive the sight of the tag motion. This approach takes a snapshot of tags positions through their backscattered strength very several read cycles, producing a sequence of strength frames. We implement the system using pure COTS RFID devices and evaluate it at various parameter choices.

Slide 28

Thanks ! Q&A