No More Potentially Dynamic Objects: Static Point Cloud Map Generation based on 3D Object Detection and Ground Projection

Video [youtube]

Download datasets

KITTI dataset (https://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=3d).

• velodyne laser data (*.bin) and clibration files (times.txt).

Verions

We tested the code on python 3.8 and Ubuntu 20.04 (ROS noetic).

Setup directories

bash setup_directories.sh

Unzip and move downloaded data into project directory

Change (~/dev/no_more_potentially_dynamic_objects) to your project directory.

cd ~/Downloads/ ## change directory to the download directory
unzip ./data_odometry_velodyne.zip
mv ./dataset/sequences/00/velodyne ~/dev/no_more_potentially_dynamic_objects/data/KITTI/origin 
unzip ./data_odometry_calib.zip
mv ./dataset/sequences/00/calib/times.txt  ~/dev/no_more_potentially_dynamic_objects/data/KITTI

Setup the environment

cd ~/dev/no_more_potentially_dynamic_objects ## change directory to the project directory
conda create -n no_more_dynamic python=3.8
conda activate no_more_dynamic
pip install -r requirements.txt

Publish point clouds

roscore

Open the different terminal (no anaconda environment is necessary. Instead, rospkg needs to be installed.)

python mini_kitti_publisher.py --dir ./data/KITTI

Object detection

We used VoxelNeXt as an object detection method in our paper. For the environment setting, we followed documents on OpenPCDet.

Additionally, we visualized bounding boxes to test the inference results. To use visualization code, we prepared corresponding pcd and pkl file.

python visualization.py

Ground segmentation

We used patchwork-plusplus as an object detection method in our paper. Follow the setup process in patchwork-plusplus

python ground_segmentation.py --dir ./data/KITTI

Ground projection

python ground_projection.py --dir ./data/KITTI

Mapping

Publish point clouds with projected objects

python mini_kitti_publisher.py --dir ./data/KITTI --velodyne_dir projected

Proceed mapping

We used SC-A-LOAM as a mapping method in our paper.

Acknowledgement

The code in this repository is based on VoxelNeXt, patchwork-plusplus, SC-A-LOAM, and mini-kitti-publisher. Thanks to the authors of those codes.

License

Copyright 2023, Soojin Woo, Donghwi Jung, Seong-Woo Kim, Autonomous Robot Intelligence Lab, Seoul National University.

This project is free software made available under the MIT License. For details see the LICENSE file.