Camera Calibration

Camera Calibration

Setup

Calibration depends upon torch, torchvision, numpy and opencv-python packages. These should have been installed when you used the installation script to install LoCoBot and PyRobot.

Run Calibration (One time)

  1. Start all the ROS services.
# Launch the arm controllers, ar_tag tracker, etc. Make sure to specify the
# correct robot base (kobuki or create).
roslaunch locobot_calibration calibrate.launch base:=kobuki
  1. Run script to collect data for calibration (should take about 10 minutes). The robot will move the arm and camera around and record images and joint angles. Keep robot arm workspace clear of obstacles.
source ~/pyenv_pyrobot/bin/activate
cd ~/low_cost_ws/src/pyrobot/robots/LoCoBot/locobot_calibration/

# By default the calibration routine will use 6 fixed poses. You can modify
# these poses as necessary in scripts/collect_calibration_data.py
python scripts/collect_calibration_data.py \
    --data_dir tmp/calibration-data/v1/ \
    --botname locobot
  1. Run script to solve for calibration.
python scripts/solve_for_calibration_params.py \
    --data_dir tmp/calibration-data/v1/ \
    --calibration_output_file ~/.robot/calibrated.json \
    --overwrite --n_iters 10001 --to_optimize camera_link
  • Calibration solver setups up a least square pixel re-projection error between the estimated location of the AR tag corners using the kinematic chain, and the detected location of AR tag corners using basic image processing.
  • Transforms listed in scripts/solve_for_calibration_params.py are optimized using gradient descent (using Adam optimizer).
  • Calibration parameters are saved to calibration_output_file.
  • Note that the solver first optimizes for translation parameters and then estimates both translation and rotation.
  • The solver prints out the pixel error. In practice, we have observed a pixel error of upto 5 pixels gives reasonable performance. If the solver is only able to converge to a larger error:
    1. Verify the data being used for calibration (data_dir will contain a directory tag_detect that will contain the images with the detected tag corners, visually scan to make sure they are accurate and don't have too many outliers),
    2. Play around with the solver parameters (learning rate, Adam parameters, robust loss parameters, etc.).
    3. Additionally, also optimize for the arm_base_link transform. This can be done using the arm_base_link to --to_optimize flag, as: --to_optimize camera_link,arm_base_link.

Publish Estimated Transforms

Running the following service will publish the estimated calibration parameters. These will get incorporated in the ROS transform tree automatically. This service must be run anytime the camera is used. If you are using locobot_control/main.launch, this will be automatically launched and you do not have to run it separately.

roslaunch locobot_calibration calibration_publish.launch
← Install software[Basic] Navigation →