Awesome
multicam_calibration - extrinsic and intrinsic calbration of cameras
Installation
Download the package:
mkdir catkin_ws
cd catkin_ws
mkdir src
cd src
git clone https://github.com/KumarRobotics/multicam_calibration.git
You will need the following packages in your ROS workspace:
git clone https://github.com/catkin/catkin_simple
git clone --recursive https://github.com/versatran01/apriltag.git
And probably libceres:
sudo apt install libceres-dev
What else you are missing you'll find out now:
cd catkin_ws
catkin config -DCMAKE_BUILD_TYPE=Release
catkin build
How to use:
First, produce the best starting guess you can come up with,
and edit it into calib/example/example_camera-initial.yaml:
cam0:
camera_model: pinhole
intrinsics: [605.054, 604.66, 641.791, 508.728]
distortion_model: equidistant
distortion_coeffs: [-0.0146915, 0.000370396, -0.00425216, 0.0015107]
resolution: [1280, 1024]
rostopic: /rig/left/image_mono
cam1:
T_cn_cnm1:
- [ 0.99999965648, -0.00013331925, 0.00081808159, -0.19946344647]
- [ 0.00013388601, 0.99999975107, -0.00069277676, -0.00005674605]
- [-0.00081798903, 0.00069288605, 0.99999942540, 0.00010022941]
- [ 0.00000000000, 0.00000000000, 0.00000000000, 1.00000000000]
camera_model: pinhole
intrinsics: [605.097, 604.321, 698.772, 573.558]
distortion_model: equidistant
distortion_coeffs: [-0.0146155, -0.00291494, -0.000681981, 0.000221855]
resolution: [1280, 1024]
rostopic: /rig/right/image_mono
Adjust the topics to match your camera sources.
You must use an aprilgrid target for calibration, layout follows Kalibr conventions and
is specified in config/aprilgrid.yaml.
Then launch the camera calibration:
roslaunch multicam_calibration calibration.launch
You can see the camera images and the detected tags overlaid with any of the ros
image visualization tools.
There is a sample perspective file in the config directory:
rqt --perspective-file=config/example.perspective
Then play your calibration bag (or do live calibration):
rosbag play falcam_rig_2018-01-09-14-28-56.bag
You should see the tags detected, and output like this on the terminal:
type is multicam_calibration/CalibrationNodelet
[ INFO] [1515674455.127216052]: added camera: cam0
[ INFO] [1515674455.130332740]: added camera: cam1
[ INFO] [1515674455.131238617]: not using approximate sync
[ INFO] [1515674455.132790610]: writing extracted corners to file corners.csv
[ INFO] [1515674458.646217104]: frame number: 10, total number of tags found: 349 336
[ INFO] [1515674459.958243084]: frame number: 20, total number of tags found: 698 686
[ INFO] [1515674461.349852261]: frame number: 30, total number of tags found: 1048 1036
... more lines here ....
[ WARN] [1515674512.667679323]: no detections found, skipping frame!
[ INFO] [1515674512.757430315]: frame number: 410, total number of tags found: 11896 13300
When you think you have enough frames collected, you can start the calibration:
rosservice call /multicam_calibration/calibration
This should give you output like this:
Num params: 2476
Num residuals: 201928
iter cost cost_change |gradient| |step| tr_ratio tr_radius ls_iter iter_time total_time
0 4.478809e+03 0.00e+00 5.32e+06 0.00e+00 0.00e+00 1.00e+04 0 2.45e-01 3.10e-01
1 1.291247e+03 3.19e+03 2.03e+05 1.46e+00 1.55e+00 3.00e+04 1 5.11e-01 8.21e-01
2 1.288842e+03 2.40e+00 6.22e+03 2.38e-01 1.04e+00 9.00e+04 1 4.56e-01 1.28e+00
3 1.288794e+03 4.79e-02 3.19e+02 3.57e-02 1.02e+00 2.70e+05 1 4.37e-01 1.71e+00
4 1.288792e+03 2.27e-03 3.73e+01 7.64e-03 1.01e+00 8.10e+05 1 4.38e-01 2.15e+00
5 1.288792e+03 2.61e-05 5.09e+00 7.20e-04 1.01e+00 2.43e+06 1 4.38e-01 2.59e+00
6 1.288792e+03 6.92e-08 5.35e-01 3.46e-05 1.03e+00 7.29e+06 1 4.37e-01 3.03e+00
Solver Summary (v 1.12.0-eigen-(3.2.92)-lapack-suitesparse-(4.4.6)-cxsparse-(3.1.4)-openmp)
Original Reduced
Parameter blocks 410 410
Parameters 2476 2476
Residual blocks 409 409
Residual 201928 201928
Minimizer TRUST_REGION
Sparse linear algebra library SUITE_SPARSE
Trust region strategy LEVENBERG_MARQUARDT
Given Used
Linear solver SPARSE_NORMAL_CHOLESKY SPARSE_NORMAL_CHOLESKY
Threads 4 4
Linear solver threads 1 1
Linear solver ordering AUTOMATIC 410
Cost:
Initial 4.478809e+03
Final 1.288792e+03
Change 3.190017e+03
Minimizer iterations 7
Successful steps 7
Unsuccessful steps 0
Time (in seconds):
Preprocessor 0.0653
Residual evaluation 0.0680
Jacobian evaluation 1.4113
Linear solver 1.5961
Minimizer 3.2011
Postprocessor 0.0000
Total 3.2663
Termination: CONVERGENCE (Function tolerance reached. |cost_change|/cost: 1.930077e-13 <= 1.000000e-12)
[ INFO] [1515674589.074056064]: writing calibration to /home/pfrommer/Documents/foo/src/multicam_calibration/calib/example/example_camera-2018-01-11-07-43-09.yaml
cam0:
camera_model: pinhole
intrinsics: [604.355, 604.153, 642.488, 508.135]
distortion_model: equidistant
distortion_coeffs: [-0.014811, -0.00110814, -0.00137418, 0.000474477]
resolution: [1280, 1024]
rostopic: /rig/left/image_mono
cam1:
T_cn_cnm1:
- [ 0.99999720028, 0.00030730438, 0.00234627487, -0.19936845450]
- [-0.00030303357, 0.99999829718, -0.00182038902, 0.00004464487]
- [-0.00234683029, 0.00181967292, 0.99999559058, 0.00029671670]
- [ 0.00000000000, 0.00000000000, 0.00000000000, 1.00000000000]
camera_model: pinhole
intrinsics: [604.364, 603.62, 698.645, 573.02]
distortion_model: equidistant
distortion_coeffs: [-0.0125438, -0.00503567, 0.00031359, 0.000546495]
resolution: [1280, 1024]
rostopic: /rig/right/image_mono
[ INFO] [1515674589.251025662]: ----------------- reprojection errors: ---------------
[ INFO] [1515674589.251045482]: total error: 0.283519 px
[ INFO] [1515674589.251053450]: avg error cam 0: 0.28266 px
[ INFO] [1515674589.251059520]: avg error cam 1: 0.284286 px
[ INFO] [1515674589.251070091]: max error: 8.84058 px at frame: 110 for cam: 1
[ INFO] [1515674589.251410620]: -------------- simple homography test ---------
[ INFO] [1515674589.331235450]: camera: 0 points: 47700 reproj err: 0.440283
[ INFO] [1515674589.331257726]: camera: 1 points: 53252 reproj err: 0.761365
In the calib/example directory you can now find the output of the calibration:
ls -1
example_camera-2018-01-11-08-24-22.yaml
example_camera-initial.yaml
example_camera-latest.yaml
Parameters
use_approximate_sync: (default: false) uses the ROS approximate sync framework to approximately synchronize image frames with different message header stamps.corners_file: if a corners file is specified, such corners file is loaded as input data when the calibration node starts up, as if these corners had been detected by feeding images to the calibration node. This allows repeating of previously done calibrations by keeping the corners file instead of all the images. Whenever points are fed into the calibration node, it writes the corners to~/.ros/corners.csv.run_calib_no_init: run calibration right after loading the corners file. This is mostly for debugging purposes.fix_intrinsics: fixes all intrinsics. Note that more fine-grained fixing of intrinsics for individual cameras can be done on the fly with ROS service calls.record_bag: was supposed to record the images that were used for calibration, but this feature is currently broken due to some ROS bug.outlier_pixel_threshold(default: -1). If specified greater than 0 will remove any detected corners that exceed the error threshold and re-run the calibration again. Note: new option, has not seen much testing yet.output_filename,latest_link_name,calib_dir, andresults_dircombined specify where to look for the initial files and where the calibration results will go. The parameterization is somewhat confusing so it's best to look at the example launch files and/or the source code.detector_type: (default:Mit) allows to switch between the MIT and theUmichversion of the apriltag implementationtag_border: (only valid if using theMitdetector, specifies the width of the black border frame of the tags, defaults to 2).
Managed calibrations
Sometimes a calibration consists of a sequence of steps, for example: first the intrinsics of each sensor, then the extrinsics of the sensors with respect to each other. This is particularly useful when image data between sensors is not synchronized.
To help with this, you can write a little python program that does
that. In fact, you just have to modify the section below in
src/example_calib_manager.py, and voila, when you trigger your
calibration manager, it will in turn run multiple calibrations via
service calls into the calibration node, each time retaining the
previous calibration's output as initial value. Here is an example
section, adjust as needed:
# first do intrinsics of cam0
set_p(FIX_INTRINSICS, "cam0", False)
set_p(FIX_EXTRINSICS, "cam0", True)
set_p(SET_ACTIVE, "cam0", True)
set_p(FIX_INTRINSICS, "cam1", True)
set_p(FIX_EXTRINSICS, "cam1", True)
set_p(SET_ACTIVE, "cam1", False)
run_cal()
# then do intrinsics of cam1
set_p(FIX_INTRINSICS, "cam0", True)
set_p(FIX_EXTRINSICS, "cam0", True)
set_p(SET_ACTIVE, "cam0", False)
set_p(FIX_INTRINSICS, "cam1", False)
set_p(FIX_EXTRINSICS, "cam1", True)
set_p(SET_ACTIVE, "cam1", True)
run_cal()
# now extrinsics between the two
set_p(FIX_INTRINSICS, "cam0", True)
set_p(FIX_EXTRINSICS, "cam0", True)
set_p(SET_ACTIVE, "cam0", True)
set_p(FIX_INTRINSICS, "cam1", True)
set_p(FIX_EXTRINSICS, "cam1", False)
set_p(SET_ACTIVE, "cam1", True)
run_cal()
Unit tests
For unit testing of the calibration code, refer to this page.