lidar_bev is a ROS package that computes the Bird's Eye View of a LiDAR pointcloud. It follows the encoding defined in BirdNet. This representation is also used by the newer BirdNet+.

This package was developed at the Intelligent Systems Laboratory, Universidad Carlos III de Madrid, by Jorge Beltrán, Carlos Guindel, Francisco M. Moreno, Irene Cortés, and Daniel Cruzado.

Where to start?

• If you want to generate BEV images for training and validation of models on the KITTI Object Benchmark, please run:

  roslaunch lidar_bev offline_bev.launch

• For online operation, a ROS node is provided. Launch files for different LiDAR devices are located in launch/.

• The files in scripts/ are mainly intended to perform the post-processing/refinement stage of the original BirdNet in order to obtain rotated boxes from the Faster R-CNN axis-aligned detections. This has been deprecated by BirdNet+.

Run configuration

Bird's Eye View parameters

There are many parameters to configure the BEV. Default values lead to results published in BirdNet and BirdNet+. These are: camera_fov, planes, h_res, v_res, low_opening, cell_size, grid_dim, grid_min_x, grid_max_x, grid_min_y, grid_max_y, max_height, min_height, num_slices, height_threshold, cell_size_height_map, grid_dim_height_map, get_ground, remove_floor.

Offline mode

• kitti_dir: Path to kitti/object folder.
• split_dir: KITTI split set. Can be training or testing.
• saving_path: folder name (or full path) to save resulting images. If no full path is provided, the folder will be created in $HOME/.ros/

Online mode (ROS node)

• lidar_tf_frame: name of the LiDAR frame
• camera_tf_frame: name of the camera frame

Inputs (Subscribed Topics)

• cloud_topic (sensor_msgs::PointCloud2): LiDAR pointcloud

Outputs (Published Topics)

• bird_view (sensor_msgs/Image): BEV image
• bird_ground (sensor_msgs/Image): Image representing the ground estimation
• ground_cloud (sensor_msgs::PointCloud2): Pointcloud to visualize the ground estimation in Rviz.

Dependencies

This repository depends on ROS. ROS Kinetic and newer versions should work. Additionally, an auxiliary script is used to compute the density channel of the BEV (max_points_map.py). To setup the dependencies, please run:

pip install progress numpy

ROS Dependencies

• cv_bridge
• image_transport
• pcl_ros
• sensor_msgs
• tf

Citation

If you find this module useful, consider citing one of our works:

The original BirdNet, where this BEV encoding was introduced.

@inproceedings{Beltran2018,
    title = {{BirdNet: a 3D Object Detection Framework from LiDAR information}},
    author = {Beltrán, Jorge and Guindel, Carlos and Moreno, Francisco Miguel and Cruzado, Daniel and García, Fernando and de la Escalera, Arturo},
    booktitle = {Proc. IEEE International Conference on Intelligent Transportation Systems (ITSC)},
    pages = {3517--3523},
    publisher = {IEEE},
    year = {2018}
}

The next generation, BirdNet+.

@misc{Barrera2020,
    title = {{BirdNet+: End-to-End 3D Object Detection in LiDAR Bird's Eye View}},
    author = {Barrera, Alejandro and Guindel, Carlos and Beltrán, Jorge and García, Fernando},
    booktitle = {arXiv:2003.04188 [cs.CV]},
    url = {http://arxiv.org/abs/2003.04188},
    year = {2020}
}