[feat] Add an optional overlap estimator

• This overlap estimator allows to evaluate the registration quality at each new scan input

• It is based on the LCP estimator implemented in SUPER4PCS random registration algorithm

• It represents the probability for one scanned point to have a neighbor in the keypoint maps.

• It is computed as an average of this probability for all scanned points. These probabilities are estimated with a gaussian like density fitted in map leaf space.

• The function can be decomposed in 3 steps :

  • Transform the input points into new frame
  • Sample uniformly if needed
  • Compute customized LCP value

• Time info on one example -> indoor data set, fixed maps 8300 points and input scan 29000 points, no sampling :

  • Without any multithreading: 5.9ms
  • With all multithreading (2 threads): 3.2ms
  • Without Transform multithreading : 3.6ms
  • Without LCP multithreading : 5.5ms
  • The LCP step is the heaviest
  • The sampling is not neglectable

• How to interpret :

  • A good overlap estimation lies in [0.7, 1]
  • A sufficient one lies in [0.4,0.7]
  • A bad one lies in [0, 0.4]

• It is particularly helpful to detect bad convergences (big jumps) because the overlap must show a gap. Warning: this is true when the map is fixed but can be false if we pass through a door while the map is built online.

• DOUBT : should we keep the uniform filtering option?

  • adv :
    • it can reduce LCP computation time
    • it allows not to focus on one points cluster overlap
  • inc :
    • it can allow some overlap gap along the SLAM process with no particular reason
    • it is long with respect to the whole process

@nicolas.cadart PTAL