1. KITTI Dataset Pre-processing
February 24, 2024 · View on GitHub
[TOC]
1. KITTI Dataset Pre-processing
1.1 Remove Dynamic Objects
1.1.1 Download and Assign
Please download the KITTI dataset from here and the Semantic KITTI dataset from here, and place them in the data_preprocess/kitti/dataset/sequences/00 directory.
1.1.2 Remove Dynamic Objects
First, create a directory named pcd within data_preprocess/kitti/dataset/sequences/00/ to save the point cloud of the removed dynamic objects.
Next, you can use the script bin2pcd_folder.py located at data_preprocess/scripts/ to remove dynamic objects and convert the point cloud data format from .bin to .pcd. You can execute the following command:
python data_preprocess/scripts/bin2pcd_folder.py \
--input_dir data_preprocess/kitti/dataset/sequences/00/velodyne \
--output_dir data_preprocess/kitti/dataset/sequences/00/pcd
The script bin2pcd_folder.py utilizes the labels from the Semantic KITTI dataset to identify and remove the point clouds associated with labeled dynamic objects.
with open(binFileName, "rb") as f:
byte = f.read(size_float * 4)
while byte:
# refer to semantic-kitti-api/config/semantic-kitti.yaml [https://github.com/PRBonn/semantic-kitti-api]
if(((label[pcd_count]>=252) and (label[pcd_count]<=259)) # remove dynamic objects
or (label[pcd_count]==0) # unlabeled
or (label[pcd_count]==1) # outlier
or (label[pcd_count]==10) # car
or (label[pcd_count]==11) # bicycle
or (label[pcd_count]==13) # bus
or (label[pcd_count]==15) # motorcycle
or (label[pcd_count]==16) # on-rails
or (label[pcd_count]==18) # truck
or (label[pcd_count]==20) # other-vehicle
or (label[pcd_count]==30) # person
or (label[pcd_count]==31) # bicyclist
or (label[pcd_count]==32) # motorcyclist
or (label[pcd_count]==99) # other-object
or (label[pcd_count]==251) # moving lidar-mos mod moving
):
byte = f.read(size_float * 4)
pcd_count=pcd_count+1
continue
x, y, z, intensity = struct.unpack("ffff", byte)
list_pcd.append([x, y, z])
byte = f.read(size_float * 4)
pcd_count=pcd_count+1
1.1.3 One Test Scene in Our Paper
For our test scene, we have selected scan order numbers 1151 to 1200 from the pcd directory located at data_preprocess/kitti/dataset/sequences/00/pcd. These scans constitute our test scene, which we have named KITTI-sequence00-1151-1200. This test scene comprises a total of 50 scans and can be found in the pcd_remove_dynamic1151_1200 directory within the same path.
1.2 Fuse Multi-Scan Point Cloud
To fuse the scans in the train set of the KITTI-sequence00-1151-1200 scene, you can use the script pointcloud_fusion.py located at data_preprocess/scripts/. This script offers various parameters for customization. Some of the parameters within the pointcloud_fusion.py script include:
Lidar_height=0.168
data_start=1150 # scan order number from data_start+1
data_end=1200
root_dir="data_preprocess/kitti/dataset/sequences/00/pcd_remove_dynamic1151_1200"
kwargs = {'root_dir': root_dir, 'data_start':data_start, 'data_end':data_end,
'range_delete_x':3, 'range_delete_y':2, 'range_delete_z':1.25,
'interest_x':20,'interest_y':20, # interest area of single scan
'pose_path':"data_preprocess/kitti/dataset/sequences/00/poses.txt",
'save_path':"data_preprocess/kitti_pre_processed/sequence00/1151_1200_view/",
'over_height':Lidar_height, 'over_low':-2
}
You can execute the point cloud fusion script using the following command:
python data_preprocess/scripts/pointcloud_fusion.py
Make sure you are in the appropriate directory when running this command and that any required input data is available.
1.3 Ground Filtering
cd data_preprocess/scripts/tool_ws
catkin_make
source devel/setup.bash
roslaunch plane_ground_filter plane_ground_filter.launch
Please note that you should set the pcd_path parameter in the plane_ground_filter.launch file located at data_preprocess/scripts/tool_ws/src/plane_fit_ground_filter/ with the absolute path to the directory. During this step we have referred to AbangLZU.
<arg name="pcd_read1" default="true" />
<group if="$(arg pcd_read1)">
<param name="pcd_path" value="/home/meng/subject/data_preprocess/kitti_pre_processed/sequence00/1151_1200_view/source.pcd" />
<node pkg="plane_ground_filter" name="read_pub_pcd" type="read_pub_pcd.py" output="screen">
</node>
</group>
The ground filtering result is as follows:
You can save the ground point cloud and non-gorund point cloud as:
cd data_preprocess/kitti_pre_processed/sequence00/1151_1200_view
mkdir sub_pointcloud
cd sub_pointcloud
# Save the ground point cloud and rename it as points_ground.pcd:
rosrun pcl_ros pointcloud_to_pcd input:=/points_ground .
# Save the non-ground point cloud and rename it as points_no_ground.pcd:
rosrun pcl_ros pointcloud_to_pcd input:=/points_no_ground .
the result:
After running the provided commands to rename the point clouds, you will have two files in the current directory:
1.4 Non-Ground Point Cloud Clustering
To perform non-ground point cloud clustering, you can create some directories to organize your results.
cd data_preprocess/kitti_pre_processed/sequence00/1151_1200_view/sub_pointcloud
mkdir no_ground_clusters
mkdir child_nerf
mkdir child_nerf2
Some details in no_ground_cluster_kitti.py:
pcd_path = "data_preprocess/kitti_pre_processed/sequence00/1151_1200_view/sub_pointcloud/points_no_ground.pcd"
pcd = o3d.io.read_point_cloud(pcd_path)
points = np.asarray(pcd.points)
points_num=points.shape[0]
folder_name = f"data_preprocess/kitti_pre_processed/sequence00/1151_1200_view/sub_pointcloud/no_ground_clusters/"
clusters,aabb_list = region_growing_segmentation(pcd, radius=0.35, min_cluster_size=15,max_cluster_size=points_num, folder_name=folder_name)
You can conduct non-ground point cloud clustering using the no_ground_cluster_kitti.py script with the following command:
python data_preprocess/scripts/no_ground_cluster_kitti.py
The script has processed the data, and the terminal output indicates that a total of 90 clusters have been generated. These clusters likely represent segmented objects or regions within the non-ground point cloud data.
(pcnerf) meng@meng:~/subject$ python data_preprocess/scripts/no_ground_cluster_kitti.py
100%|████████████████████████████████████████████████████| 1094303/1094303 [06:41<00:00, 2725.18it/s]
uncluster point number: 396
pcd_unclusters_file_path: data_preprocess/kitti_pre_processed/sequence00/1151_1200_view/sub_pointcloud/no_ground_clusters/unclusters.pcd
len(clusters): 90
len(aabb_list): 90
sucess clusters number: 90
sucess clusters point number: 1093907
pcd_success_clusters_file_path: data_preprocess/kitti_pre_processed/sequence00/1151_1200_view/sub_pointcloud/no_ground_clusters/success_clusters.pcd
1.5 Obtain AABBs for Child NeRFs
We moved the 90 non-ground clusters from data_preprocess/kitti_pre_processed/sequence00/1151_1200_view/sub_pointcloud/no_ground_clusters and the points_ground.pcd into data_preprocess/kitti_pre_processed/sequence00/1151_1200_view/sub_pointcloud/child_nerf2, resulting in a total of 91 clusters. These clusters have been sequentially renamed from 1 to 91.
You can use the script view_pcd_aabb1.py to visualize the clusters and bounding boxes with the following command:
python data_preprocess/scripts/view_pcd_aabb1.py
This command will execute the script, allowing you to view the clusters and their associated bounding boxes.
You can further partition the clusters using the split_child_nerf_xyz.py script with adjustable intervals, as demonstrated in the code snippet below:
xy_threshold = 1 # Can be adjusted
z_threshold = 1 # Can be adjusted
split_pointcloud2(pcd_np, xy_threshold, z_threshold, output_path_prefix, points_num, i+1)
You can execute the script using the following command:
python data_preprocess/scripts/split_child_nerf_xyz.py
You can use the script view_pcd_aabb2.py to visualize the results. Here's the command to execute the script:
python data_preprocess/scripts/view_pcd_aabb2.py
