RosbagToSemanticKitti : automatic conversion from a rosbag to the SemanticKitti format (LiDAR part)

This package was made to convert a Rosbag with LiDAR frames (PointCloud2) into the architecture of the SemanticKitti Dataset, using user-friendly python scripts. This repository has also a tool to perform automatic labeling. It calculates the angles of the normals on the estimated surface described by each point, and compares them to a threshold.

Install all the requirements

This repository was tested with Python 3.8 and ROS Noetic with the sensor_msgs/PointCloud2 ROS message type. So, make sure you have it installed !

Make sure to have all the requirements for this repository. If you have pip installed, you can try :

pip install -r requirements.txt

Conversion of the point clouds

The script convert_pointcloud reads a given rosbag and write the corresponding .bin file for each frame of the LiDAR in the right folder.

• We have the following options:
  • -b, --bagfile [String] : Path to the bagfile
  • -d, --dataset [String]: Path to the dataset
  • -t, --topic [String]: Name of the topic of the Point Cloud (PointCloud2 format)
  • --id [String]: Name of the sequence (default '00')
  • -n, --name [String]: Name of the folder of the point cloud files (default 'velodyne')

These indications are also available by using -h, it will display all the above informations. Note that if the name of the sequence already exists, it will continue to add new files to it without overwriting.

Automatic labeling using normal orientation threshold

The script label_normals proceeds automatic labeling for each cloud point created with the previous script and write them in a .label file. The id of each point is 0.

We have developed this algorithm for labeling point clouds:

1. Estimate the normal on each point considering a number of neighboors to estimate the local surface
2. For each normal, compares its angle with a reference vector (here, the relative vertical)
3. If the angle is above the define threshold, the point is labeled as obstacle (2) otherwise, the point is labeled as free space (1)
4. If the intensity or the coordinates of the point are 0, the point is labeled as void (0)

• We have the following options:
  • -d, --dataset [String]: Path to the dataset
  • --id [String]: Name of the sequence (default '00')
  • -b, --bin [String]: Name of the bin folder (set previously)
  • -n, --name [String]: Name of the folder of the labels files (default 'labels')
  • -a, --angle [Float]: Threshold angle in degrees
  • --neighboor [Int]: Number of neighboors used in normal estimation

These indications are also available by using -h, it will display all the above informations

Generate poses

For each sequence, a file named poses.txt has to be generated. This, can be done by using (according to the SemanticKitti website) a surfel-based SLAM approach (SuMa).

Calibration file

Originally, in the SemanticKitti Dataset, different sensors (LiDAR and cameras) have been used. So, for each sequences, there is a file named calib.txt that contains all the projection matrices. This repository contains a default file (identity matrix for each sensor), and can be used for example if you only have one sensor and/or no calibration file.

Example

In the folder example/, you can try the execution of the different scripts by building an example dataset from a given rosbag in this folder. To do so, make sure you have the required permissions (chmod +x example.sh) and execute the bash script named example.sh. The only file that is missing is the poses.txt because it requires the execution of the SuMa script as mentionned before.

Little reminder : if you want to know the different information about the rosbag, open a terminal and type :

roscore

And in another terminal :

rosbag info bag_example.bag