cuCONERUSH

cuCONERUSH is a GPU-first ROS 2 package for real-time LiDAR cones detection. The runtime package name is cuda_cone_rush.

The processing pipeline stays on CUDA buffers across stages:

PointCloud2 -> optional XYZ filter -> optional ground segmentation -> optional voxel clustering -> cone markers

TODO

[cuda_cone_rush_node-1] [WARN] [1775749308.140774720] [cuda_cone_rush_node]: Not enough points for clustering (0 < 1)
[cuda_cone_rush_node-1] terminate called after throwing an instance of 'thrust::THRUST_200500_520_NS::system::system_error'
[cuda_cone_rush_node-1]   what():  parallel_for failed: cudaErrorInvalidDevice: invalid device ordinal
[ERROR] [cuda_cone_rush_node-1]: process has died [pid 8233, exit code -6, cmd '/home/orin/ros2_dev/install/cuda_cone_rush/lib/cuda_cone_rush/cuda_cone_rush_node --ros-args -r __node:=cuda_cone_rush_node --params-file /home/orin/ros2_dev/install/cuda_cone_rush/share/cuda_cone_rush/config/config.yaml'].

Tested on

Lidar Points	Board	avg exec (ms)
128K	Orin Nano Super	10ms
128K	Orin Dev 2048 cuda	5ms
128K	RTX 4060 Mobile 3072 cuda	1.1ms

What Is Included

• CUDA passthrough filtering (cuda_filtering)
• CUDA RANSAC ground segmentation (cuda_segmentation)
• CUDA voxel clustering + cone dimension filtering (cuda_clustering + cluster_filtering)
• Two point cloud conversion backends:
  • CUDA backend (src/utils/cuda_pointcloud_converter.cu)
  • CPU backend (src/utils/pointcloud_converter.cpp)
• Optional performance logging to ROS topic (LOGGER_PUB)

Requirements

• ROS 2 Humble
• CUDA Toolkit
• PCL
• ament_cmake

Build

From your workspace root:

cd ~/ros2_ws
colcon build --packages-select cuda_cone_rush
source install/setup.bash

CMake Options

These options are available in CMakeLists.txt:

Option	Default	Description
ENABLE_VERBOSE	OFF	Prints processing timings to stdout
LATENCY_TESTING	OFF	Print latency information for each BARQ callback
USE_CUDA_POINTCLOUD_CONVERTER	ON	Uses CUDA converter instead of CPU converter
USE_PINNED_MEMORY	ON on x86, OFF otherwise	Enables pinned host memory for host/device transfers
NSIGHT_SDK	OFF	Set an execution limit for NVIDIA Nsight Systems profiling

Example:

colcon build --packages-select cuda_cone_rush --cmake-args -DENABLE_VERBOSE=ON -LATENCY_TESTING=ON

Run

cd ~/ros2_ws
source install/setup.bash
ros2 launch cuda_cone_rush cudaconerush.launch.py

Configuration

Parameters are loaded from config/config.yaml under /cuda_cone_rush_node.

Topics

• input_topic
• filtered_topic
• segmented_topic
• cluster_topic
• frame_id

Clustering

• minClusterSize
• maxClusterSize
• voxelX
• voxelY
• voxelZ
• countThreshold

Cluster Dimension Filter

• clusterMaxX, clusterMaxY, clusterMaxZ
• clusterMinX, clusterMinY, clusterMinZ
• maxHeight

Passthrough Filter Bounds

• downFilterLimitX, upFilterLimitX
• downFilterLimitY, upFilterLimitY
• downFilterLimitZ, upFilterLimitZ

Segmentation

• distanceThreshold
• maxSegmentationDistance
• maxIterations
• probability

Pipeline Toggles and Timing Guard

• filter
• publishFilteredPc
• segment
• publishSegmentedPc
• clustering
• publishCluster
• limitWarning_ms

ROS 2 Interface

Subscriptions:

• input_topic (sensor_msgs/msg/PointCloud2)

Publications:

• filtered_topic (sensor_msgs/msg/PointCloud2) when publishFilteredPc=true
• segmented_topic (sensor_msgs/msg/PointCloud2) when publishSegmentedPc=true
• cluster_topic (visualization_msgs/msg/Marker) when publishCluster=true
• /logger/clustering/time (std_msgs/msg/Float64) when compiled with LOGGER_PUB=ON

Package Layout

src/
    main.cpp
    controller_node.cu
    filtering/cuda_filtering.cu
    segmentation/cuda_segmentation.cu
    clustering/cuda_clustering.cu
    clustering/cluster_filtering/dimension_filter.cpp
    utils/cuda_pointcloud_converter.cu
    utils/pointcloud_converter.cpp

include/cuda_cone_rush/
    controller_node.hpp
    filtering/
    segmentation/
    clustering/
    utils/

License

Apache-2.0