1. A Force-Directed Approach for Offline
GPS Trajectory Map Matching
Dr Efstratios Rappos
HEIG-VD, Switzerland
9 November 2018
ACM SIGSPATIAL, Seattle, USA

2. Background – what is Map-matching?
• Placing a trajectory obtained from GPS sensors onto a real map
• Need two things:
• the trajectory (sequence of points)
• the map
• Online versus offline:
• Real-time (online) case: we can only use info about the trajectory up to now
(focus on ‘speed’)
• Offline case: the future is also available (focus on ‘accuracy’)
• In this work we considered the offline case

3. Traditional methods
• Routing-based (using routing algorithms for matching)
• Probabilistic – Hidden Markov Chains
• Similarity based (combining data from many trajectories)
• Speed versus accuracy (ACM SIGSPATIAL 2012 competition)
• Matching can be very easy or very hard!

4. A force directed approach
• Force directed methods are common on graph / network visualization
• Video:

5. Force directed for Map-matching
• “Have the road attract the path” : novel idea
• In every road of the map, an ‘electric current’ passes though.
• Each GPS point is attracted or repelled by the road.
• Trajectory can extend or contract as necessary, within reason.
• Forces decrease with distance and over time.
• Points are allowed to move slightly under the total forces, producing a
new trajectory ‘closer to reality’, and easier to work with.
• => Trajectories converge towards the road

7. Computational experimentation
• Maps from OpenStreetMap
• Data for Taxis in Rome:
Advantage: very good dataset, because
• High road density, non-grid-like
• GPS points recorded every 10+ secs (not too often)
• Thousands of trajectories
Disadvantage
• Incompleteness in the Map data (taxis can drive to many more places than
OpenStreetMap says)

8. Map limitations

9. Computational results
• Compared the map matching produced by GraphHopper
• on the original trajectories
• with the trajectories modified according to the Force-Directed algorithm
• Two evaluation metrics found in the literature for the case of the
absence of ground truth
• Length index (comparison of trajectory lengths): consistently better, up to 5%
improvement
• Average lateral error: an average of 16% improvement
• It works no matter what the map matching algorithm is!
• Just ‘correct’ the GPS points by perturbing them slightly (pre-
processing) and improve the accuracy.

11. Video

12. Ongoing work
• Test with other datasets
• Test with other matching algorithms
• Calibrate / optimize the parameters of the algorithm
• Consider cases with constellations, flyovers, multi-lane matching
• Use the timestamp information