![● Increasing frequency of map updates
○ OSM: ~ 1.5 million new nodes/day
● Map-making is
○ Very expensive
○ Tedious and manual
● Map data is
○ Valuable
○ Becoming proprietary
[Source: http://wiki.openstreetmap.org/wiki/File:Osmdbstats7A.png]
Maps are not static… and are not cheap!](https://image.slidesharecdn.com/sdmkharitaslides-180906071148/85/Kharita-Robust-Road-Map-Inference-Through-Network-Alignment-of-Trajectories-3-320.jpg)
Kharita: Robust Road Map Inference Through Network Alignment of Trajectories
- 1. Robust Road Map Inference Through Network Alignment of Trajectories May 3rd, 2018 Rade Stanojevic, Sofiane Abbar, Saravanan Thirumuruganathan, Sanjay Chawla @ Qatar Computing Research Institute, HBKU Fethi Filali @ Qatar Mobility Innovations Center Ahid Aleima @ Land Transport Planning Department, MoTC, Qatar
- 2. Map Inference from GPS Traces GPS Traces Road Network
- 3. ● Increasing frequency of map updates ○ OSM: ~ 1.5 million new nodes/day ● Map-making is ○ Very expensive ○ Tedious and manual ● Map data is ○ Valuable ○ Becoming proprietary [Source: http://wiki.openstreetmap.org/wiki/File:Osmdbstats7A.png] Maps are not static… and are not cheap!
- 4. High industrial value, autonomous vehicles
- 5. Why work on this problem? Why Qatar? ● Despite massive investment, maps have errors: fundamental challenges to overcome ● Fast Growing Cities (Doha) ○ Extensive construction: frequent diversions, new roads, road blocks, etc. ○ Commercial maps (e.g., Google) often out of date ○ Working with local stakeholders to update maps (e.g., MoTC, QMIC, Karwa)
- 6. Why work on this problem? Why Qatar?
- 7. GPS-only: Prior work ● Clustering Schemes (Edelkamp et al.) ● Trace merging from Microsoft (Cao et al.) ● KDE schemes from U. Chicago (Biagioni et al.)
- 8. Problem Definition Given a collection of GPS Trajectories <T>, infer a directed and weighted graph <G> representing the underlying road network. ● T = {tr1,... trn} ● tri = {x1, … xmi} ● xi = <latitude, longitude, timestamp, angle, speed>
- 9. Kharita Overview ● Kharita infers the topology (directed graph) of the underlying road network based solely on GPS data ● Kharita models the problem as a “multi network alignment problem - MNAP” where each GPS trace is one network ● MNAP is modeled as a nonlinear extension of the Facility Location Problem
- 10. MNAP for Map Inference Directed Graph ● Nodes: intersections/junctions ● Edges: road segmens.
- 11. Kharita: Non linear extension of FLP Standard Facility Location Objective Term to infer the edges Sparsification S.T: i j k l O_ik S_jl
- 12. Kharita: Non linear extension of FLP For “small size” problems we can solve the optimization problem using a Lagrangian relaxation heuristic Standard Facility Location Objective Term to infer the edges Sparsification S1. K-means clustering S2. Edge Inference S3. Spanners
- 13. Kharita: Offline Algorithm ● Distance Metric ○ Combine lat, lon, and angle into one distance metric (penalize angles difference.) ○ L_i: (lat, lon); 𝛼_i: angle, 𝜃: heading penalty ○ E.g., An angle difference of 180 should be equivalent to a geometric distance of 50 meters
- 14. Kharita: Offline Algorithm ● Densification ○ Create fictional points to deal with varying sampling rates ● Clustering ○ One cluster every x meters ● Edge Inference ○ Geometry of clusters ○ Trajectories ● Graph Sparsification ○ Remove spurious edges GPS Traces Clustering Edge Inference Graph spanner
- 15. Kharita*: Online Algorithm ● Streaming ○ Process trajectory edges one at a time ● Trace merging ○ Combine clustering and edge inference into one step ● Adapt a variant of streaming graph spanner
- 16. Kharita*: Online Algorithm ● Streaming ○ Process trajectory edges one at a time ● Trace merging ○ Combine clustering and edge inference into one step ● Adapt a variant of streaming graph spanner
- 17. Evaluation ● Data ○ QMIC data from Doha (lat, lon, ts, angle, speed) ○ UIC data from Chicago (lat, lon, ts) ● SOTA ○ Edelkamp et al. ’03 (k-means) ○ Cao et al. SIGSPATIAL’09 (Trace merging) ○ Biagioni et al. KDD’12 (KDE /Hybrid) ○ Chen et al. KDD’16 (Meanshift / Hybrid) ● Evaluation metric ○ Holes and marbles ○ GEO and TOPO Ground Truth Map Inferred Map
- 19. Evaluation: Qualitative GPS Data Kharita
- 21. Concluding Remarks ● Keep it simple ● Always check the quality of your data ○ Good data collection rules can lead to very accurate maps. E.g., ■ Generate points every x seconds, or ■ Every time the angle changes significantly, or ■ Every time the speed changes significantly, etc. ○ Use as much data as you can ■ If you don’t have angle (or/and speed) information, infer them. ● Online map update is not easy ○ Revisit the network to fix errors ● Code on github: https://github.com/vipyoung/kharita
- 22. QCRI is hiring for its new Center for Artificial Intelligence (Q-CAI). Come and talk to us! FYI Sofiane Abbar (sabbar@qf.org.qa) Sanjay Chawla (schawla@qf.org.qa)
- 23. Thank You!