Multi-modal Probablistic 室内定位 on a Smartphone

摘

• (GPS)
  • robust loc on smartphone outdoor
• indoor, no system
  • to achieving a similar level of ubiquity

 

• a multi-modal positioning
  • targeted at smartphone
  • aims to get the best out of each of its constituent
    modalities
• combine
  • (BLE) beacons
  • RTT enabled WiFi
  • smartphone’s (IMU)
  • provide a cheap robust loc that
• unlike fingerprinting, no pre-training

 

• a probabilistic algorithm based on a (CRF)
• incorporate sparse visual information to
  • improve the accuracy of our system
  • using pose estimation from pre-scanned visual landmarks,
  • to calibrate the system online

 

• achieves an accuracy of 2 meters outperforming other distance-based loc
• compare our approach with (UWB) system
  • do not match the performance of UWB
  • our system is cheap, smartphone compatible
  • provides satisfactory performance for many applications

 

• Smartphone loc,
• multi-modal systems,
• RTT-based ranging
• crf

I. INTRODUCTION

• existing systems
  • expensive dedicated infrastructure
  • training/calibration
  • not robust to changes of the environment
  • require the user to actively scan the environment
  • not smartphone-compatible

 

• ultra-wideband (UWB) signals can provide accurate
  over a wide range of distances with a quick and stable system setup [14]
• requires dedicated infrastructure
• no : become smartphone compatible

 

• Other solutions require only a smartphone and no infrastructure

• cameras combined with (IMU) can be used to scan the environment and, using simultaneous localization and mapping (SLAM), infer an accurate
  3D map and the user’s location with centimeter-level precision [33]

• not always convenient for the user to actively scan the surroundings;

• ideal 室内定位 would be passive

  • no participation of the user

• visual systems also require

  • the environment to
  • possess uniquely identifiable visual features
  • not always
  • exhibition rooms,
  • lecture halls, and sterile environments such as hospitals

• visual systems

  • break down when the environment is changed,
  • the recorded maps to not match the reality

 

• to create a passive loc

• a third category of approaches use smartphone-compatible measurements

• (RF) signals such as (BLE) and WiFi

• WiFi base stations are prevalent

  • BLE beacons inexpensive

• indoor , RF signals are heavily affected

  • by multi-path, shadowing and fading [13]

• synchronization issues

  • unknown latency times
  • unknown variations in antenna characteristics
  • induce important measurement errors [26]

• best algorithms leveraging such signals

  • use a “fingerprinting” phase,
  • in which the signal’s characteristics are extracted at a finite set of positions in the region of interest

• These fingerprints are able to capture stationary non-line-of-sight (NLOS) and multipath effects

• online loc phase,the recorded signal is matched with the fingerprint database

  • after possible combination with IMU data and map constraints—
  • a position estimate is inferred [10], [29], [34]

 

• an indoor localization
  • smartphone compatible
  • easy to install
  • robust to environmental changes
• combines
  • BLE beacons, WiFi hotspots
  • the smartphone’s IMU data
  • and visual fingerprints recorded by its camera.
• visual information is sparse in time and space
• assume that certain visual landmarks
  • artworks in a museum
  • emergency exit signs
  • navigation maps
  • do not change over a time and these landmarks can be “scanned” by the user at a handful of time points
• At these moments, obtain centimeter-level loc,
  • allows us to calibrate offsets in the other measurements.

 

• CRF to combine the different measurement modalities
• allows to account for the expected relative accuracy of the different systems
• obtain a distribution
  • rather than a single estimate of the user’s position

 

• provides robust loc
  • not break down when measurements or entire modalities are missing or unreliable
• passive for the majority of time
• mean localization errors of around two meters in a challenging indoor , without any fingerprinting or prior calibration

2. RELATED WORK