230 likes | 614 Views
FreeLoc : Calibration-Free Crowdsourced Indoor Localization. Sungwon Yang, Pralav Dessai , Mansi Verma and Mario Gerla UCLA. Outline. Introduction Fingerprint value extraction Localization algorithm Evaluation. Introduction.
E N D
FreeLoc: Calibration-Free Crowdsourced Indoor Localization Sungwon Yang, PralavDessai, MansiVermaand Mario Gerla UCLA Neight @ NSlab Study group
Outline • Introduction • Fingerprint value extraction • Localization algorithm • Evaluation Neight @ NSlab Study group
Introduction • Investigate 3 major technical issues in crowd sourced indoor localization system: • No dedicated surveyor. Can’t afford long-enough time for survey and Can’t sacrifice their device resources • No constraint on type & number of device. • There are no designated fingerprint collection points. Different user can upload their own fingerprint with same location label. • Contributions: • Present a method that extracts a reliable single fingerprint value per AP from the short-duration RSS measurements • Proposed novel indoor localization method, requires no calibration among heterogeneous devices, resolves the multiple surveyor problem • Evaluate system performance Neight @ NSlab Study group
System overview Send measured RSSI and request location info. • Multiple-surveyor-Multiple-user System • Every one is contributor & user • Fast radio map building & update • Similar system exists, but still some challenges not being addressed in the related work A,B upload Fingerprint data with location label Neight @ NSlab Study group
System Challenges • RSS Measurement for short duration • Multi-path fading in indoor environment cause RSSIto fluctuate overtime • To construction a robust and accurate radio map, more RSSI samples is better • Update map / large area is time consuming • Short-time measurement is necessary • Device Diversity • Different designed hardware ( Wi-Fi chipset, antenna,…etc ), RSSI varies even though collect at the same location • Multiple Measurements for one location in crowd sourced system • Different surveyor might reply different RSSI fingerprint even though they are in the same location area. • Multiple fingerprints for a location is not effecient Neight @ NSlab Study group
Outline • Introduction • Fingerprint value extraction • Localization algorithm • Evaluation Neight @ NSlab Study group
Fingerprint value extraction • AP response rate • AP were not recorded in some fraction of the entire Wi-Fi scanning duration • Their preliminary result: • RSSI > -70dbm provides over 90% response rate • -70dbm < RSSI < -85dbm provides 50% response rate • RSSI < -90dbm provides very poor response rate • Given lower weight to weak RSSI, discount the AP response rate for fingerprint information Neight @ NSlab Study group
Fingerprint value extraction • RSS variance over time • RSSI value observation result in their testbed • Top figure : collect RSSI for 1 HR • Middle/Bottom : collect for 1 minute • Collect frequency: 0.5-1Hz, depend on different device • Related works often suggests using the mean value of RSSI or using Gaussian distribution model • Fig.(a) an example, the RSSI histograms are strongly left-skewed. Gaussian model can’t fit well. • Also, mean value is not always the best idea Fig.(a) an example, mean value work well Fig.(b) an example, long time & short time variation could degrades the localization accuracy. Neight @ NSlab Study group
Extraction Method • Observation Findings: • The most-recorded RSSI in the case of the short duration measurements is very close to the most recorded RSSI in long-duration cases • fpValueis the fingerprint value for an AP • RSSpeakis the RSS value with highest frequency • The width of the range being averaged is set by and • Select stronger RSS value as the fpValueif more than one RSS value has the same frequency in a histogram • However, it’s difficult to adjust and and RSSpeakmove slightly left or right each timedepend on environment factors Neight @ NSlab Study group
Extraction Method Modified • Modified Fingerprint model • Use one width w and set it enough large • Euclidean distances between Fpvalue from one-hour measurement and one-minute measurement with respect to log scale • Averaging 50 measurements and more than10 AP recorded in each measurement and find w Neight @ NSlab Study group
Outline • Introduction • Fingerprint value extraction • Localization algorithm • Evaluation Neight @ NSlab Study group
Localization Algorithm BSSID vector, Fingerprint of location lx Keyi is the BSSID with ith strongest RSSI • Relative RSS comparison Surveyors Users Neight @ NSlab Study group
Localization Algorithm Let us see the example… Neight @ NSlab Study group
Localization Algorithm Location result would be in 101 8pts • Relative RSS comparison Surveyors Users 1pts Neight @ NSlab Study group
Localization Algorithm Location result would be in 101 9pts • Relative RSS comparison Surveyors Users 2pts Neight @ NSlab Study group
Localization Algorithm If no high rank key match, label location as unknown High rank key • Relative RSS comparison Surveyors Users Neight @ NSlab Study group
Heterogeneous Devices • Radio map work well, even though heterogeneous devices involved. • Due to not use absolute RSS value, but utilize relationship among RSSI • Therelieves the degradation of localization accuracy. AP not detected Neight @ NSlab Study group
Multiple Surveyors • More than one user can upload their own fingerprints • Maintain only one fingerprint • Update fingerprint become possible, by merge fingerprint Neight @ NSlab Study group
Evaluation adjacent of point 1.5m Corridor width 2.5m • Environment Setup • 70 different locations at the engineering building in university • Fingerprint comprised information • Timestamp • BSSID (MAC address) • RSSI • Four different devices • Motorola Bionic, HTC Nexus One, Samsung GalaxyS and GalaxyS2 • Two main scenario result would be show in this work adjacent of point 6m Neight @ NSlab Study group
Pairwise Devise Evaluation Overall, best delta value is 12 In laboratory, best delta value is around 12, Cross device error<2m • Find out whether the proposed method of building fingerprint and using it for indoor localization works well with heterogeneous devices • Find out the optimal δ value, to be used for subsequent experiments • Collect data over 3 days Neight @ NSlab Study group In 3rd Floor, best delta value is about 9, Cross device error<4m
Different device fingerprint not affect localization accuracy Merge Fingerprint mechanism might help to increase localization accuracy Impact of Device Heterogeneity • Wi-Fi fingerprinting data for each location was taken from multiple devices and data from all other mobile phone devices In 3rd Floor In laboratory Neight @ NSlab Study group
Impact of Multiple surveyors • Constructed the fingerprint map for a particular room using heterogeneous devices placed at different parts (levels) of the room. • The user requesting for location information was assumed to be standing at the center of the room. • Every level had three devices, that were different from the user’s device. The higher level would farer from the center. • limits the error in accuracy to less than 3 meters Neight @ NSlab Study group
Discussion • Magic point: About utilizing the relationship not value for localization • Future work: • Filtering erroneous fingerprint data is essential in crowd-sourced systems • Since the entire system is based on participation of untrained normal users • Outdated fingerprint data may significantly degrade the localization accuracy • Merge algorithm would failed… Neight @ NSlab Study group