If you are interested only in the CNN part, you can find the models and the pretrained networks in the res folder.

The code has been tested with ROS Lunar and PyTorch 0.4.1. You can find an inclusive list of dependencies in the repo. They will probably contain useless dependencies.

First of all, clone the repository in your ROS environment. Some dependencies are managed by conda, install the required packages with

conda install --file requirements_conda.txt

Then, use pip for the remaining dependencies

pip install -r requirements_pip.txt

Compile the project with a standard catkin_make in your ROS environment root.

Two launch configuration are available in the launch folder. After compiling the code, they should be available with (example)

roslaunch ld_lsi debug

The difference between the debug configuration and the standard one is that with the debug view, OpenCV is used to display the output of the CNN and of the clustering algorithm. If you want to use other stream names as inputs for the ROS nodes, please edit the corresponding fields in the launch files.

If you found this code useful for your research, please cite

@article{pizzati2019enhanced,
  title={Enhanced free space detection in multiple lanes based on single CNN with scene identification},
  author={Pizzati, Fabio and Garc{\'\i}a, Fernando},
  journal={arXiv preprint arXiv:1905.00941},
  year={2019}
}