* see /Vision/Data/kitti_velodyne

* Before commit "[feat] Extract all keypoints ", one point over
neighborhood_size was processed to be a potential keypoint, this was
reducing the computation time of keypoints extraction

* Now, all points are processed so the computation time can be increased
in case of dense clouds (32/64/128 Lidar sensors)

* So a parameter is added to downsample the input clouds when extracting
the keypoints to speed up the process.

* The downsampling is random

* /!\ This downsampling only speeds up the keypoints extraction step,
the other steps are mostly driven by the maximum number of extracted keypoints

* If the frame is not 360°, with new space gap the points at the field
of view extremities could be extracted which is not expected. This
problem affects more cases now that first and last points are reenabled.

* Therefore min/max azimuth angles are added to invalidate some points due
to robot/vehicle/pedestrian occlusion or simply because the user wants to
reduce the field of view

* Remove saliency parameter

* Add space gap parameter : : minimum number of points that must be 
missed
to consider a feature

* Add max keypoints number and voxel grid resolution to wrapping

* + add interface to log only keyframes

* Add tools to convert isometries to ros messages
* Refact gps callback
* Refact optimize graph command
* Output optimized trajectory
* Remove doubled g2o file
* Remove experimental params
* Add interface to calibrate GPS/Lidar
* Broadcast gps offset
* The Calibration path are not outputed anymore, g2o file can be used to visualize results

* Add command to launch Pose graph with landmarks (this command takes an 
optional file name argument)

* Add graph parameters to config file

* Adapt launch file to add a private parameter to rapidly access the 
tag_detections topic

* remove all Transform occurances -> add wrappers to complete functions

* The Tukey loss is used to reject tag outliers based on the saturation
distance

* This allows to reject tags info if the tag has been moved for example
or if a false detection appears

* Add an interface to parameterize the landmark constraint building

* Add a callback to receive April tag output, to transform it with
camera pose information and to store it into the SLAM object. The
covariance is rotated too, using CERES.

* Add a function to load an initial file containing the absolute pose
information of the tags that can be encountered. The file can have a
header and contain the tag ID, its 6D pose and the relative covariance
(43 elements) in this order.

* Add public argument to tag topic in launch file

* Add a new external_sensor group to gather sensor info

* Adapt documentation to new external sensor structure

* The third eigen value is already evaluated with the max mse allowed.

* The two others will just be used in the plane case, to not use too
linear neighborhood as planes.

* The well-known planarity and linearity values could not be used
because of the very low number of points used to build a model.

* The ratio of the minimum plane chosen is 1/5 (which leads to an eigen
values ratio of 1/25)

It is updated to 1.5m (in ROS: 1m indoor and 1.5m outdoor)
In most situations, there is no object in this radius, so this changes 
nothing.
However, in pedestrians or indoor datasets, it might useful to use 
points really close to the lidar.

- Some important features were missing from the indoor config file
- Minor change in outdoor config file

* Add a computation time in the confidence output
* Add a config file for indoor applications and specify the existing
config file as outdoor
* Update documentation with those changes
* Add comment to detail steps order

* Update mode was set to 3 (value can be 0, 1 or 2)

* The minimum number of frames having reached one voxel can be set to reject moving objects (-> parameter min_frames_per_voxel).
* If 0, the behavior is the same as before and all voxels are extracted.
* The submaps are outputed in a new topic.
* rviz config is updated to display these new outputs.

* Add voxel_grid/decaying_threshold parameter

* 3 parameters are added :
   * voxel_grid/sampling_mode_planes
   * voxel_grid/sampling_mode_edges
   * voxel_grid/sampling_mode_blobs

* They allow to choose which point to keep in a voxel when downsampling the keypoints' maps

* The Lidar SLAM angular estimation is not smooth enough to estimate 
accurately local velocity.

* The velocity estimation is smoothed, using a temporal window 
parameterizable in seconds.

* In PV, loggingTimeout param follows WindowTime param as it is not 
settable from UI

* In ROS, the user has to choose a loggingTimeout value g.t. WindowTime 
param manually. If not, a warning is raised and the estimation is based 
on the current stored poses.

Rename matching parameters for clearer usage.

MinNbrMatchedKeypoints -> MinNbMatchedKeypoints
MaxDistanceForICPMatching -> MaxNeighborsDistance

LineDistanceNbrNeighbors -> EdgeNbNeighbors
MinimumLineNeighborRejection -> EdgeMinNbNeighbors
LineDistancefactor -> EdgePcaFactor
MaxLineDistance -> EdgeMaxModelError

PlaneDistanceNbrNeighbors -> PlaneNbNeighbors
PlaneDistancefactor1 -> PlanePcaFactor1
PlaneDistancefactor2 -> PlanePcaFactor2
MaxPlaneDistance -> PlaneMaxModelError

BlobDistanceNbrNeighbors -> BlobNbNeighbors

Delete unused KeypointsMatcher::Params::MinNbrMatchedKeypoints

This allows setting this parameter from ParaView UI.

This param value was lowered from 25 yo 10 neighbors.

To build a model of the neighborhood, a minimum number of points is needed:
- 2 for a line model
- 3 for a plane model
- 4 for an ellipsoid model

A check is added to verify that enough neighbors are requested to build the target model.
Otherwise, a MatchStatus::BAD_MODEL_PARAMETRIZATION error code is returned.

This also prevents the matching procedure from crashing if no neighbors are requested.

Use OverlapSamplingRatio = 0 to disable overlap.
Set default to OverlapSamplingRatio = 0, automatically disabling overlap computation to save time.
Update wrappings to change this behavior.

VoxelGrid sampling is quite slow, and leads to smaller overlap values.
Now, the idea is to use only a subset of points to compute overlap on.
The user can provide a ratio of points to use, between 0 and 1.

* Before there were 4 values (linear part + roll, pitch, yaw parts)

* Now there are 1 linear part and 1 angular part

* Uniformization constraint/threshold/limit -> limit

* Thresholds for velocity and acceleration