This package is an implementation of litamin2 point cloud matching algorithm.

ROS PCL CERES Eigen3

mkdir -p litamin_ws/src/ && cd litamin_ws/src/
git clone https://github.com/bzdfzfer/litamin2
cd ~/litamin_ws/ && catkin build

source devel/setup.bash
rosrun litamin2 litamin2_align ~/litamin_ws/src/litamin2/data/251370668.pcd ~/litamin_ws/src/litamin2/data/251371071.pcd

.

White --> Source cloud.

Green --> Target cloud.

Blue --> LM optimization alignment.

Red --> Ceres optimization alignment.

• [GT] Two scan transformation matrix:

    0.999941    0.0108432 -0.000635437     0.485657
  -0.0108468     0.999924  -0.00587782      0.10642
 0.000571654   0.00588436     0.999983   -0.0131581
           0            0            0            1

Voxel resolution is set to 0.5m.

• LM based optimization with zero lambda.

• speed: repeative run of 100 times, average 32.7774[msec].
• accuracy:

align result: 
   0.999867   0.0162508 -0.00157465    0.491412
 -0.0162532    0.999867 -0.00147708    0.117267
 0.00155044  0.00150248    0.999998  -0.0306296
          0           0           0           1

• Ceres based solution.

• speed: repeative run of 100 times, average 132ms.
• accuracy:

align result: 
   0.999875   0.0157213 -0.00137114    0.465903
 -0.0157243    0.999874 -0.00218937    0.114884
 0.00133655  0.00221065    0.999997  -0.0321361
          0           0           0           1

Voxel resolution is set to 3m.

• LM optimization.

• speed: 100 times run average time: 38.8906[msec]
• accuracy:

align result: 
   0.999903   0.0138753 -0.00154203    0.489225
 -0.0138811    0.999896 -0.00385291    0.116165
 0.00148841  0.00387394    0.999991  -0.0299192
          0           0           0           1

• Ceres:

• speed: 100 times run average time: 461.8[msec]
• accuracy:

align result: 
   0.999902   0.0138793  0.00193534    0.455404
 -0.0138774    0.999903 -0.00103081    0.110234
-0.00194946  0.00100385    0.999998  -0.0399503
          0           0           0           1

• Newton optimization

• accuracy:

TO BE Debuged ....

Zoom in details.

• Yokozuka M, Koide K, Oishi S, et al. LiTAMIN2: Ultra Light LiDAR-based SLAM using Geometric Approximation applied with KL-Divergence, ICRA2021. litamin2 paper link.
• Kenji Koide, Masashi Yokozuka, Shuji Oishi, and Atsuhiko Banno, Voxelized GICP for fast and accurate 3D point cloud registration, ICRA2021. fast-gicp paper link.
• Official implementation of Fast-GICP. fast-gicp implementation.
• Another implementation of Fast-GICP using Ceres. fastgicp re-implementation.

* Current implementation only uses ICP cost, the covariance rotation cost function is not successfully added. (due to instable convex optimization of trace terms. ) * Find Hessian matrix of E_{ICP} and E_{COV} and implement Newton method mentioned in the paper.

The Hessian of ICP cost is added in current implenentation, but the contribution on accuracy improvement is negligible. This proves that jacobian approximatin is good enough for hessian computation.

Someone interested in the derivation process of Hessian matrix and Jacobian matrix, please refer to "litamin2_derivatives.pdf".

If someone knows how to compute the derivative of trace with respect to lie so(3), please teach my in the issue.

I have derived the jacobian and hessian matrix of COVARIANCE cost part in "litamin2_derivatives.pdf".

The implementation of this part is coded at "litamin2point2voxelnewton_impl.hpp".

!!!!(Sad story) New added jacobian and hessian did not test passed !!!！

The KITTI odometry dataset is used to run this algorithm, part of seq 00 is shown below.

To run this test, one can use odometry dataset:

rosrun litamin2 test_kitti_align path_to_your_kitti_odometry_dataset KITTI output_path

Or run with raw dataset:

rosrun litamin2 test_kitti_align path_to_your_kitti_raw_dataset KITTI_raw output_path