LAMI 1. User Oriented Trajectory Similarity Search 2. Calibration-free Localization using Relative Distance Estimations 3. From GPS Traces to a Routable Road Map Radu Mariescu-Istodor
Similarity Search Naive solution is to compare Every pair of points (10 8 ) From reference route and Every route in DB (10 5 ) Average route length ~1000 points haversines ~ 30 minutes
Limiting the space R-tree indexed points => easy to conduct range checks Naive solution is to compare Every pair of points (10 8 ) From reference route and Every route in DB (10 5 ) the routes that have at least one point in one of the squares
Reducing range check calculations Grouping several squares into a bounding rectangle
How to group? Dead Space Limiting the Dead Space as much as possible
How to limit the Dead Space?
Longest Common Subsequence Heaviest Common Subsequence O(N 2 )
User Oriented Similarity Heaviest Common Subsequence User defined Regions
Improve speed of HCSS Heaviest Grouped Subsequence HCSS ≤ HGSS 1. HGSS = HGSS = HGSS = HGSS = 301 ………………. HCSS ?
Purpose How to calculate the position of mobile devices that do not posses a GPS sensor
Centroid and Fingerprinting
How it works? Cell1=50% Cell2=80% Cell3=60% GPS signature
How it works? Cell1=50% Cell2=80% Cell3=60% Cell1=80% Cell3=20% Cell5=30% Cell6=20% Common Cells = 2 Uncommon Cells = 3 Spearman = ? Feature : (Experimentally deduced)
Regression Formula (Experimentally deduced) Fitted from GPS phones Features ={Common, Uncommon, Spearman}
Estimating locations
Objective: Road Network from GPS Tracks
Merging nearby trajectories Gravity
Merging nearby trajectories GravitySpring Resultant force applied in small iterations until change insignificant
Traces of opposite directions Method so far Desired output:
Repelling force Sign tells if same direction or not ≠ ?
Solution
Banding issue expecting Gaussian Distribution
Matching the Gaussians expecting Gaussian Distribution Finding centroids gives number of bands