lidar_align

A simple method for finding the extrinsic calibration between a 3D lidar and a 6-dof pose sensor

Note: Accurate results require highly non-planar motions, this makes the technique poorly suited for calibrating sensors mounted to cars.

The method makes use of the property that pointclouds from lidars appear more 'crisp' when the calibration is correct. It does this as follows:

1. A transformation between the lidar and pose sensor is set.
2. The poses are used in combination with the above transformation to fuse all the lidar points into a single pointcloud.
3. The sum of the distance between each point and its nearest neighbor is found. This process is repeated in an optimization that attempts to find the transformation that minimizes this distance.

Installation

To install lidar_align, please install ROS Indigo, ROS Kinetic or ROS Melodic.

The following additional system dependencies are also required:

sudo apt-get install libnlopt-dev

Input Transformations

The final calibrations quality is strongly correlated with the quality of the transformation source and the range of motion observed. To ensure an accurate calibration the dataset should encompass a large range of rotations and translations. Motion that is approximately planner (for example a car driving down a street) does not provide any information about the system in the direction perpendicular to the plane, which will cause the optimizer to give incorrect estimates in this direction.

Estimation proceedure

For most systems the node can be run without tuning the parameters. By default two optimizations are performed, a rough angle only global optimization followed by a local 6-dof refinement.

The node will load all messages of type sensor_msgs/PointCloud2 from the given ROS bag for use as the lidar scans to process. The poses can either be given in the same bag file as geometry_msgs/TransformStamped messages or in a separate CSV file that follows the format of Maplab.

Visualization and Results

The node will output it's current estimated transform while running. To view this your launchfile must set output="screen" in the <node/> section. See the given launchfile for an example.

Once the optimization finishes the transformation parameters will be printed to the console. An example output is as follows:

Active Transformation Vector (x,y,z,rx,ry,rz) from the Pose Sensor Frame to  the Lidar Frame:
[-0.0608575, -0.0758112, 0.27089, 0.00371254, 0.00872398, 1.60227]

Active Transformation Matrix from the Pose Sensor Frame to  the Lidar Frame:
-0.0314953  -0.999473  0.0078319 -0.0608575
  0.999499 -0.0314702 0.00330021 -0.0758112
 -0.003052 0.00793192   0.999964    0.27089
         0          0          0          1

Active Translation Vector (x,y,z) from the Pose Sensor Frame to  the Lidar Frame:
[-0.0608575, -0.0758112, 0.27089]

Active Hamiltonen Quaternion (w,x,y,z) the Pose Sensor Frame to  the Lidar Frame:
[0.69588, 0.00166397, 0.00391012, 0.718145]

Time offset that must be added to lidar timestamps in seconds:
0.00594481

ROS Static TF Publisher: <node pkg="tf" type="static_transform_publisher" name="pose_lidar_broadcaster" args="-0.0608575 -0.0758112 0.27089 0.00166397 0.00391012 0.718145 0.69588 POSE_FRAME LIDAR_FRAME 100" />

If the path has been set the results will also be saved to a text file.

As a method of evaluating the quality of the alignment, if the needed path is set all points used for alignment will be projected into a single pointcloud and saved as a ply. An example of such a pointcloud can be seen below.

CSV format

Note that Maplab has two CSV exporters. This file-format is the same as produced by exportPosesVelocitiesAndBiasesToCsv but differs from the output of exportVerticesAndTracksToCsv

Parameters

Scan Parameters

IO Parameters

Alinger Parameters