1 / 37

Using GPS to learn significant locations and predict movement across multiple users

Using GPS to learn significant locations and predict movement across multiple users. Daniel Ashbrook , Thad Starner College Of Computing, Georgia Institute of Technology, From Personal and Ubiquitous Computing, vol 7, no 5, 2003. Preface. Early work in this area

topper
Download Presentation

Using GPS to learn significant locations and predict movement across multiple users

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Using GPS to learn significant locations and predict movementacross multiple users Daniel Ashbrook, Thad Starner College Of Computing, Georgia Institute of Technology, From Personal and Ubiquitous Computing, vol 7, no 5, 2003

  2. Preface • Early work in this area • Very simple method with 324 citation!! • Both authors’ magnum opus • And the papers cite it also have large number of citation

  3. Outline • Introduction • Previous work • The pilot study • The Zürich study • Application • Future work • Conclusion • Comments

  4. Outline • Introduction • Previous work • The pilot study • The Zürich study • Application • Future work • Conclusion • Comments

  5. Introduction • Wearable computers as intelligent agents assist the user in a variety of tasks • Location is the most common context to determine the users’ tasks • We present a system that automatically find the significant locations • Create user models to predict movement

  6. Outline • Introduction • Previous work • The pilot study • The Zürich study • Application • Future work • Conclusion • Comments

  7. Previous Work • Sparacino used infrared beacons to create individual models of museum visitors • Liu and Maguire describe a generalized network architecture that incorporated prediction with the goal of supporting mobile computing • Above using fixed sensors, but systems using GPS must determine which locations are significant

  8. Previous Work • Researches using GPS • Wolf used stopping time to mark the starting and ending points of trip • Marmasse and Schmandt used the loss of GPS signal to detect buildings

  9. In This Research • The goal: Construct a system to record and model an individual’s travel and predict on different scale • Answer the query like “Where is Daniel most likely to go after work?” • Conducted two studies. • In 2001, a pilot study with single user in four months • In 2002, six users in seven months

  10. Outline • Introduction • Previous work • The pilot study • The Zürich study • Application • Future work • Conclusion • Comments

  11. The Pilot Study - data • One user for a period of four months • GPS receive rate is once per second • Valid signal and moving at one mile per hour at least • The user traveled in and around Atlanta

  12. The Pilot Study - methodology • Find significant places • Clustering places into locations • Learning sublocations • Prediction Definition: place = gps points which are significant location = clustered places

  13. Finding significant places • Latitude and longitude are useless • “Home” and “Work” are meaningful • The logical way to finds significant points is to look at • Where the user spends her/his time • Significant locations will be inside buildings(no GPS signal) • We define a “place” with an interval of time t between it and the previous point

  14. Finding significant places • We need to decide what value of time t • There is no clear point on the graph to choose • Use 10 minutes as stopping time

  15. Clustering places into locations • Because of the erroneous GPS measurements, the logger won’t record exactly the same point even the user stops • Use K-means to cluster the places • place list and a radius until no place remained • Every cluster denotes a “location”, and is assigned a unique ID

  16. Clustering places into locations • So we need to decide the radius • Too small or too large may cause problem • We run our clustering algorithm several times with varying radius • And find the “knee” point • Decide next n and threshold? Knee point: For each point on the graph, we find the average of it and the next n points on the right. If the current point exceeds the average by some threshold, we use it as the knee point.

  17. Learning sublocations • We subsume smaller-scale paths • From city-wide scale to campus-wide scale • Taking the points within each “location” and running the same clustering algorithm • If a knee exists, it forms a sublocation

  18. Prediction • We substitute for each place the ID of the location it belongs to • Markov model is created for each location with transition probability • Node is location, edge is transition probability • First order and second order… nth order • Quantity of second order is small…

  19. It’s first order Markov model A Edge - transition B C Compare the path’s relative frequency to the probability that the path was taken by chance (Monte Carlo simulation)

  20. Outline • Introduction • Previous work • The pilot study • The Zürich study • Application • Future work • Conclusion • Comments

  21. The Zürich study – data • Conducted a second study in Zürich, Switzerland with multiple users • Six users during seven months • Unfortunately, one user broke the cabling for his unit and was unable to collect any data at the beginning • Total 800,000 data points

  22. The Zürich study – methodology • Find places using time t • But now we register a place when signal is lost • Cluster places to locations • Markov model

  23. Place 1 Place 2 Fewer places

  24. The Zürich study – evaluation • We present two results • The correlation between the names assigned to locations by users • Even in a different environment, the prediction generates consistent results

  25. Naming across users • Asked each user to give names to each location we found • To see if the locations we found are common places to the users • Ex. If all user give location A the same name • Of the five users, three had 11 locations within Zürich, one had 9 and one had 6 (it just express the time period they stay effect the number of locations)

  26. Prediction Monte Carlo simulation result User 1 User 2

  27. Outline • Introduction • Previous work • The pilot study • The Zürich study • Application • Future work • Conclusion • Comments

  28. Application • Single-user application : • To-do list with location reminder • Detect situation • Multi-user application (share user’s model) • Answer social query like “Will I see Daniel today?” • Schedule a meeting (time & location) for several people • Make a serendipitous meeting with friends • Do me a favor

  29. Outline • Introduction • Previous work • The pilot study • The Zürich study • Application • Future work • Conclusion • Comments

  30. Future Work • Support time prediction • Our model takes long time to update • Weighting update • Online learning • Suggest names for a location • To-do application • Combine two similar users’ location

  31. Outline • Introduction • Previous work • The pilot study • The Zürich study • Application • Future work • Conclusion • Comments

  32. Conclusion • We develop an algorithm to extract significant places and locations • Predictive model demonstrated patterns of movement that occurred much more frequently than chance

  33. Outline • Introduction • Previous work • The pilot study • The Zürich study • Application • Future work • Conclusion • Comments

  34. Comments • Concept is simple • Few experiment with poor evaluation • The method they proposed becomes a standard scenario for location-based activity recognition • The early bird catches the worm • Give many application • I can’t see any color

  35. Recently research • Most of them are devoted to extract significant locations using either GPS or other sensor then predict the movement or the activity • Once we have the place and activity information, we can answer social query like “Who will I meet in lab on Friday morning?” • But no more detail about users’ character • Maybe we can classify the user into category and guess who have the same interest or in the same college or … • It seems like an application rather than research…

More Related