ctyfang / amz-targetless-cal

Target-less Camera-LiDAR Calibration

Description

Targetless Calibration between Camera and LiDAR using

1. Manually selected correspondence
2. Automated correspondence detection combining Gaussian Mixture Model (GMM), Chamfer Matching and Mutual Information

Implementation assumes a KITTI dataset folder structure like the following:

dataset
├── calibration
│   ├── calib_cam_to_cam.txt
│   ├── calib_imu_to_velo.txt
│   └── calib_velo_to_cam.txt
├── image_00
│   └── data
│       ├── 0000000001.png
│       └── ...
└── velodyne_points
    └── data
        ├── 0000000001.bin
        └── ...

Requirement

We recommend using pipenv and the Pipfile.lock file to install necessary dependencies in a virtual environment.

pipenv install --ignore-pipfile
pipenv shell

Usage

python calibrate.py --dir <path to kitti directory> --calibration_method <either "automatic" or "manual">
python calibrate.py -d <path to kitti directory> -cm <either "automatic" or "manual">

Example:

python calibrate.py --dir data/0928-KITTI-dataset/ --calibration_method automatic

For manual mode, click the feature in camera image and intensity image, confirm the selection with y, and finish selection with q

The calibration result is then displayed in the consol after the algorithm has converged. Example:

### Optimization completed ###
xyz (m): [ 1.20684236 -1.21968689  1.20059145]
Rotation Vector: [-0.02272116 -0.07837192 -0.34919117]

amz-targetless-cal's Issues

The result order

Is there a mistake here
print('xyz (m): ' + str(np.squeeze(self.tau)[:3])) print('Rotation Vector: ' + str(np.squeeze(self.tau)[3:]))
I think the order of R and T is reversed, it‘s a small problem.