Overview

This library is forked from the StereoVision library publicly available on GitHub. It's purpose is to find the intrinsic and extrinsic camera parameters between stereo cameras. Note that rectification is only possible for cameras with the same image size (i.e. two identical cameras). The output of the primary script is a set of numpy (.npy) files with the extrinsic and intrinsic camera parameters to be used in various computer vision applications. This output is currently used for the Multimode Stitching repository.

Installation

To install, follow these steps:

1. Clone the repository.
2. Clone the environment using conda:

   conda env create -f environment.yml

3. source activate StereoCamera
4. python setup.py build
5. python setup.py install

Checking Your Setup

I've included a folder with test images to make sure everything is working correctly in /test_images. I got the test images from here.

Running the Calibration

To calibrate a stereo camera setup, images must be saved in the format _left_abc123.jpg and _right_abc123.jpg, where abc123 is any alpha-numeric suffix (that can be fed into python's sort() function). Each pair of images (i.e. right and left image of the same scene) must have the same suffix. All images to be used for calibration should be saved in the same folder without any other contents.

To run a stereo camera calibration after the library is installed, use the command

time calibrate_cameras --rows nrows --columns ncols --square-size squareSize --show-chessboards --validate-results inputImageFolder outputResultsFolder

Where:

• nrows, ncols = number of interior rows and columns (don't count the outermost squares on the grid)
• squareSize = size of square in centimeters
• --show-chessboards = (optional) displays detected pattern corners for each image (activate window and press any key to advance to next image)
• --validate-results = (optional) uses the calculated homography matrix to project detected chessboard corners from the right frame to the left and displays all right images overlaid on the corresponding left images

Example command for test images folder:

time calibrate_cameras --rows 6 --columns 9 --square-size 5 --validate-results test_images/3/ test_images/test_set_3_results/

More information about the original library can be found as published by the original author on his blog.

Note on Thermal Imagery

For raw thermal image files, a special process is required to make them usable in the calibration.

Size

First, the calibration (specifically, chessboard corner finding) works best if the thermal images are resized using image magick to a minimum size. The images have only been tested down to a minimum size of 960x768px. If the raw images are below this, use the following command to resize them to meet this minimum size. For example, if the images are .png files with starting size 320x256px, you would use:

magick mogrify -resize 300% *.png

Preprocessing

First, some setup is required:

1. Clone the environment using conda. A separate environment is required since the cv2 denoising function runs too slowly in the python 2 environment required for the camera calibration:

conda env create -f environment_thermal_tiff_resize.yml

2. source activate ThermalConvert

With the environment set, run the following command:

python thermal_preprocessor.py --folder source_thermal_folder --suffix processed_file_suffix --display_images --do_not_save

source_thermal_folder is the full path of the folder containing the thermal images to be preprocessed. Note that all images in this folder will be processed.

processed_file_suffix (optional) is a suffix to be added to the end of the file names before processing to create the processed image file names.

--display_images (optional) is a flag to display each image as it is processed

--do_not_save (optional) is a flag to indicate that the processed images should not be saved.

The processed images will be saved in the same folder as the source images with the same name (and added suffix, if applicable) as .jpg files.