A new interface to FLANN's multiple randomized trees for high-dimensional (feature) searches

[andersgb1]

This PR addresses PCL's missing interface to a proper search structure for high-dimensional data, as discussed here:

http://www.pcl-users.org/KdTree-radiusSearch-giving-bad-output-td3806817.html

CHANGES IN THE COMMIT:

• A new nested class KdTreeIndexMultiCreator inside FlannSearch for creating FLANN's multiple randomized tree index, suitable for approximate high-dimensional searches. The number of trees defaults to 4 (as in FLANN), but can be changed within the creator class constructor.
• A new member variable checks_ for controlling the search accuracy during all instantiations of flann::SearchParams performed inside FlannSearch. The default value is set to the same as the one used in FLANN (32) upon construction of FlannSearch.
• Made the typedefs inside FlannSearch public, because they are needed from outside.

TODO:

• Automatic use of DefaultFeatureRepresentation<FeatureT> instead of the current DefaultPointRepresentation<FeatureT> when passing KdTreeMultiIndexCreator to the constructor of FlannSearch. Even though DefaultPointRepresentation already correctly vectorizes the feature part of features, it is semantically bad.
• Automatic initialization of the multi-index instead of the single one when FlannSearch is templated on a feature (dimension > 10).
• As far as I can see, Histogram<N> is the only feature that does not have a DefaultPointRepresentation, I guess because it does not have a fixed size like e.g. SHOT352 has. It would be nice to provide point registrations of this class for N = {1, ..., N_max}, making DefaultFeatureRepresentation complete for all possible built-in PCL features. Otherwise, the user has to call POINT_CLOUD_REGISTER_POINT_STRUCT() when he wants to use a Histogram.
• Unit test and tutorial, if needed?

[jspricke]

Hi Anders, thanks for the contribution. What is your proposal for the TODO, do you want to add it into this pull request first, or should we merge it in the current state?
@amuetzel, as you are the original author, can you have a quick look, if you are ok with the patch? Thanks!

[andersgb1]

Hi @jspricke

The TODOs (except for the last point which is trivial) are still very much unknowns from my side, and I would need help with decisions to make a PR. So my intention was to leave the current PR as is, and then open up a new issue with the first three points, if the current PR is accepted.