Improve performance when passing through tunnels

[giorgiovisani]

Dear developers,
I do appreciate your hard work to produce this nice and handy package.

However, I'm struggling when it comes to tuning the hyperparameters of the matcher. In particular, I am testing it with GPS locations recorded during a motorway trip, in which we crossed few tunnels. Tunnels seem to be difficult to handle and I wanted to ask your opinion on whether it is best to:

• Increase the obs_noise (to get a larger sigma coefficient, which does not penalize too much my correct path in terms of emission probability)
• Increase the max_lattice_width to keep track of a larger number of possible paths in the HMM (hopefully keeping the path through the tunnel)
• Increase non_emitting_length_factor: my intuition is "in a tunnel there have to be many subsequent non emitting states, which I do not want to penalize".
  If you have any insights on any other method to improve your algorithm performance in the presence of a tunnel, I'm very interested in your opinion.

Also, I did not get exactly how the non_emitting_length_factor parameter is exploited in the code and how it should be tuned (eg. increasing it we obtain less or more penalty for a sequence of non-emitting states? And what is the neutral value to get no penalty neither reward for the sequence of non-emitting states against a normal sequence?)

[wannesm]

The method is robust if the outliers are spread out. But if there is a longer sequence of scattered measurements you should consider applying a Kalman filter first to denoise the signal.

For example using

LeuvenMapMatching/leuvenmapmatching/util/kalman.py

Line 19 in 9ca9f0b

def smooth_path(path, dt=1, obs_noise=1e-4, loc_var=1e-4, vel_var=1e-6, kf=None,

[giorgiovisani]

Thank you Wannes for your suggestion.
Indeed it works much better when I first apply the Kalman filter.
Keep up the good work :)