This weights IMU based on the time between nodes in the
pose graph optimization.
When moving slowly or stopping, IMU weights are reduced.
This improves quality in these cases. The parameters are
changed to approximately get the same behavior while
moving as before for the examples.
Signed-off-by: Wolfgang Hess <whess@lyft.com>
This integrates the imu based extrapolator implementation, adds lua configuration and conversions from lua to proto. Note that the parameters of the imu based extrapolator are not tuned yet and the default is still the constant-velocity extrapolator.
Signed-off-by: mschworer <mschworer@lyft.com>
This integrates the extrapolator interface into the 3D trajectory builder. The construction is now done within the interface and local trajectory builder 3D just makes use of the interface. No functional changes.
Signed-off-by: mschworer <mschworer@lyft.com>
This PR adds the possibility to use the TolerantLoss function for INS in pose graph optimization. This is currently switched off, and so there's not functional change.
The ceres TolerantLoss function (see [description](http://ceres-solver.org/nnls_modeling.html), and [implementation](https://github.com/ceres-solver/ceres-solver/blob/master/internal/ceres/loss_function.cc)) has the following property
- for large values of x it approaches a quadratic loss ("null loss") with the specified weight (fixed_frame_pose_translation_weight)
- for small values of x it approaches a quadratic loss with a smaller weight
- there's a crossover at some value x_c. The function is convex everywhere.
This removes the hardcoded value of `collate_by_trajectory` in the cloud
map builder server and permits to configure `collate_by_trajectory`
option with lua.
This commit add support of specifying the option to collate data by
trajectory introduced in #828 in the lua configuration files.
The value is set to false as when a boolean option is not specified it
is initialized to false.
Previously, two config values were spelled "bandwith".
The assumption is that these values are not used in other
repositories so the rename is harmless.
This allows Cartographer to use normal Application Default Credentials
(ADC) instead of a hand-generated format.
Google-internally, we'll only use ADC going forward. At your option, I
can keep the old code for the token files around if you think it would
be generally useful. Many OSS programs follow the same approach.
Adds a parameter for the currently hard coded `kSubmapsToKeep` in `pure_localization` mode.
The new parameter `max_submaps_with_pure_localization` allows to specify the history size of submaps in `pure_localization` mode.
Closes#1116
Add the parameter `OptimizationProblemOptions.use_online_imu_extrinsincs_in_3d`. Set it to true by default to not change existing behavior.
Using online IMU extrinsics is not always desirable, and particularly during localization, we have determined that it is preferable to turn this feature off.
- Introduce `RangeDataInserterInterface` as common interface for all range data inserters
- Introduce a minimal `GridInterface` as a base for a common interface for 2D and 3D grids
- Rename `RangeDataInserter2D` to `ProbabilityGridRangeDataInserter2D`
- Move grid generation logic from `Submap2D` to `ActiveSubmaps2D`
- Update proto and configuration structure to mirror the code structure - backwards compatibility is maintained
- Step towards [RFC 0019](https://github.com/googlecartographer/rfcs/blob/master/text/0019-probability-grid-and-submap2d-restructuring.md)
"scan" is better named "node" if it refers to a node in global SLAM.
"penalty factor" is renamed "weight" for consistency with other similar options.
"scans_per_accumulation" is renamed "num_accumulated_range_data" to match the code and since the accumulated data is not called "scan".
This PR introduces a new option that specifies the number of seconds
after which global matcher searches are performed if no recent global
constraints have been found between the submap's and the node's
trajectory.
This replaces the scaled covariances derived from the Ceres
scan matcher by directly configurable weights. Using covariances
did not provide the expected benefit, and replacing the scaling
matrix by two values will allow a faster evaluation of the cost
function in the future.
Wolfgang Hess
Martin Schwörer
Susanne Pielawa
damienrg
Andre Gaschler
Kevin Daun
Steve Wolter
Christoph Schütte
Kevin Daun
Alexander Belyaev
Alexander Reimann
Arno Fleck
Michael Grupp
Wolfgang Hess
jie
Holger Rapp
Moritz Münst
Damon Kohler
Yutaka Takaoka