The purpose of this package is to detect individual objects in pointcloud data. Points are grouped into clusters based on proximity and published as detected objects.

NOTE: A new version of this package is available in autoware.auto.

1. Pointcloud preprocessing

   • Points closer than a distance of remove_points_upto meters are removed from the cloud.
   • Points are then downsampled if the downsample_cloud parameter is set to true.
   • The pointcloud is trimmed to remove points based on height thresholds (clip_min_height and clip_max_height).
   • Points are further trimmed based on their y position to either side of the vehicle if keep_lanes is set to true. The bounds are defined by keep_lane_left_distance and keep_lane_right_distance.
   • A RANSAC-based algorithm is then used to determine a ground plane and remove any points belonging to the ground. This is activated by the remove_ground parameter.
   • The pointcloud is further filtered using Difference-of-Normals to remove any points that belong to a smooth surface. This is activated by the use_diffnormals parameter.

2. Pointcloud Clustering

   • The preprocessed pointcloud is then clustered using Euclidean Cluster Extraction, the cluster tolerance is defined by the clustering_distance parameter. This is the only part of the node that provides the option to use the GPU (activated by the use_gpu parameter).
   • Resulting clusters are then checked against neighboring clusters and any clusters which are less than cluster_merge_threshold apart are combined into a single cluster.
   • Rectangluar bounding boxes and polygonal bounds are then fit to the cluster pointclouds.

Voxel-based Downsampling
Pointcloud Surface Normal Estimation
Difference of Normals Segmentation
Euclidean Cluster Extraction

• points_raw (sensor_msgs/PointCloud2)
  Input pointcloud from lidar sensor.

• detection/lidar_detector/cloud_clusters (autoware_msgs/CloudClusterArray)
  Array of cloud clusters.
• detection/lidar_detector/objects (autoware_msgs/DetectedObjectArray)
  Array of all detected objects.
• cluster_centroids (autoware_msgs/Centroids)
  Centroids of the clusters.
• points_lanes (sensor_msgs/PointCloud2)
  Pointcloud with all preprocessing performed except Difference-of-Normals filtering.
• points_cluster (sensor_msgs/PointCloud2)
  Pointcloud colored according to cluster.
• points_ground (sensor_msgs/PointCloud2)
  Pointcloud of only ground points.

See the yaml file in the config folder for all ROS parameters and their descriptions