Cone Detection

Cluster Detection and Filtering

Cluster Detection is vital for the perception pipeline which takes a point cloud of objects (meaning the ground has been removed) and searches the cloud for clusters. This algorithm is completed within the cone_detection node, since it identifies clusters, then filters down those clusters to only ones of cones.

DBSCAN (Density-Based Spatial Clustering of Applications with Noise) was used over other clustering algorithms because it's a density-based clustering algorithm, it can discover clusters of arbitrary shapes, it doesn't need the number of clusters to be pre-specified, and it's robust to noise and outliers, all of which are beneficial to our use-case.

Overview

File Structure: cone_detection (Note that this is shared with the greater LiDAR perception pipeline.

• src/cone_detection.py

Tech Stack:

• pcl: This is the main PCL (Point Cloud Library).
• pcl_ros: A wrapper for interfacing ROS messages and PCL.
• ros2_numpy: Converts PointCloud2 message to np array for DBSCAN
• pcl_msgs
• sensor_msgs
• vision_msgs
• sklearn.cluster: imports DBSCAN for the cluster detection

Process

This cone detection process uses the following steps

• Subscription to the objects point cloud on the topic /lidar/pcl/objects

• Removes null values

• Removes values that are beyond a max and min distance from the LiDAR

• Runs cluster detection

• Removes non-cone clusters

• Publishes cone locations in message type Cone to /lidar/pcl/cones