Paper - No need to war-drive: Unsupervised indoor localization

  • Metadata:
    • author: He Wang, Souvik Sen, Ahmed Elgohary, Moustafa Farid, Moustafa Youssef, Romit Choudhury
    • title: No need to war-drive: Unsupervised indoor localization
    • year: 20122012

  • Essay:
    • While outdoor localization on phones via GPS has been around since early 2000, indoor localization techniques have yet to reach the desired accuracy for real-world deployment.
    • Even though sensors like the gyroscope and compass of modern-day smartphones can trace movements from a starting location (“dead reckoning”), this method suffers from increasing errors over time.
    • One approach to fix this would be to frequently reset the localization of the user with certain known landmarks and only use dead reckoning to capture movement between those. While you could calibrate a system to do that relying on wireless APs, this paper proposes UnLoc, a localization system solely based on mobile sensing. They assume/show that there are areas in indoor locations (stairs, elevators) with distinctive sensory fingerprints (user movement there, magnetic fields) and use classifiers to update the users' position based on the data the phone senses.
    • They use predetermined seed-landmarks/fingerprints (learned beforehand) and dynamically develop additional “organic” landmarks from the data collected by users. This allows them to theoretically start with only one seed landmark, forgoing the need for floorplans altogether.
    • They finally show that their system can reach satisfactory accuracy.
    • I enjoyed how detailed the explanation of their system and thought process while developing the concept was. They lay out their assumptions, how they tested those and how they relate to former ideas while providing supporting data for each of those. Instead of just presenting a finished idea, the paper gives insight into the development process of such systems, which is incredibly useful for me as a CS student. (Also, the bit about identifying stairs via the directionAwvalking speed correlation was quite interesting.)
    • I am, however, skeptical about how they evaluated their final approach. Seeing that their method of doing things is significantly different from formerly used techniques, they should have tried to provide as much data as possible to prove the effectiveness of the system.
    • Three test subjects in two different locations seem lacking in this regard, especially seeing how thoroughly they tested other aspects while developing the system. Most importantly, however, think they should have chosen more varied testing locations. Universities and shopping malls lend themselves pretty well to their approach, as they indeed have varying sensing fingerprints in different locations. One could imagine that parking lots, libraries and bigger convenience stores (an application the paper specifically talked about in the introduction) could pose bigger challenges when establishing organic landmarks. And while I may be wrong in my assumption that these locations would cause bigger troubles, it would have been great to see their system working in different environments to eliminate possible doubts.