iDiary: From GPS Signals to a Text-Searchable Diary Den Feldman, Andrew Sugaya, Cynthia Sung, Daniela Rus CSAIL LAB, MIT Sensys '13 2014.11.11
Goals of the paper Provides "text-searchable" database of locations and activities by taking signals generated from GPS streams
Scenarios Possible scenarios For a user, For a groups of users, What are the restaurants I visited during the last weekend? What are the roads I drove on today? For a groups of users, Where did I meet with X during the last weekend? When did Y and X meet for the first time?
Challenges Translating GPS data into useful human-understandable information (a) The data is huge 1Gbyte of data per day for a single smartphone (b) The user interface of existing GPS applications target map-navigation (c) There were very little studies, both in theory and practice
Related work Systems for analyzing location data collecting and storing mobile sensor data ex) determine transportation modes and environmental impact ex) show a list of previously visited locations Textual description of activity recognizing human activities from raw data streams ex) determine user's activity based on location information and accelerometer data manually labeled activities Semantic Compression Compressing GPS data Coresets are frequently used Text Search coresets.
iDiary Bob's semantic map Bob은 주중에 매일 집에서 직장에 간다. 주말에는 집 근처 공원에 간다. 직장에 가기 위해 기차나 자가용을 탄다. 가끔 직장에 가는 길에 아이를 유치원에 내려준다. 가끔 직장에 가는 길에 주유소에 거쳐 간다. 가끔 새로운 곳에 가기도 하지만 주로 집, 직장, 유치원, 공원, 주유소에 방문한다.
GPS data storage Collecting the signal Energy saving Bob installs the application on his mobile device(iPhone) the application collects his GPS points(traces and signals) the application transmits the to lab's server(done in the background) Energy saving save battery by switching off the GPS when the user is stationary for ten minutes switched back on once the user moves, which is detected using wi-fi and/or cell towers
GPS data storage Compressing the signal a lot of redundant information can be removed when people move from place to place, their trajectories can usually be described using linear route approximate the trajectory by a small sequence of k segments(coreset) Trajectory clustering on coresets n GPS points k routes m motion patterns
Coresets n=11, {p1,p2, .. ,p11}, k=5
Textual Description of Activity Two Database tables 2nd table 1st table begin time end time route code direction average speed route code begin point end point (latitude, longitude) (latitude, longitude) ->for each linear route (k개) ->for each motion pattern(m개)
Textual Description of Activity From numeric tables to text turns each route code into human-readable and searchable text reverse geo-coding ex) Starbucks, Anam street.. use Google Maps with latitude and longitude coordinates From text to activities ex) query "sandwich" -> return "Subway" Use LSA(Latent Semantic Analysis) each activity is a linear combination of words
User Interface
User Interface
User Interface
User Interface Query search Ex) restaurants?
Experimental results System setup 6 users using smartphone applications for a period of a year(9GB data)
Experimental results datasets
Experimental result Relevancy