autoware_ground_segmentation_cuda

Purpose

The autoware_ground_segmentation algorithms have been thoroughly tested with Autoware. However, due to latency and high computational cost when processing large pointcloud, the input pointcloud range has been limited by the crop_box_filter based on the ego-vehicle’s base_link. This can cause unwanted object loss, especially before a sloped road.

Recently, GPU and CUDA-supported libraries such as cuda_blackboard and accelerated versions of autoware_pointcloud_preprocessor have been implemented. These can be leveraged to improve the performance of ground segmentation filter algorithms using CUDA/GPU.

This package reimplements the current scan_ground_filter of the ground_segmentation package to reduce latency and avoid the bottleneck caused by processing a large number of point clouds.

Inner-workings / Algorithm

The detailed algorithm is available in scan-ground-filter.md.

Parameters

{{ json_to_markdown(“perception/autoware_ground_segmentation_cuda/schema/cuda_scan_ground_segmentation_filter.schema.json”) }}

Changelog for package autoware_ground_segmentation_cuda

0.53.0 (2026-09-29)

• Merge remote-tracking branch 'origin/main' into prepare-0.53.0-changelog
• fix(autoware_ground_segmentation_cuda): reject point clouds with an unexpected point_step (#13314) The filter kernels reinterpret_cast the cloud to a 16-byte PointTypeStruct while the host sizes its output buffers from point_step. Nothing validated the layout, so a wider point type was read at the wrong stride, and the two halves disagreed on the point size, with no error. Co-authored-by: Claude Opus 5 (1M context) <<noreply@anthropic.com>>
• refactor(perception): move node design files into each package (#13104) Co-authored-by: Claude Opus 5 (1M context) <<noreply@anthropic.com>>
• Contributors: Max Schmeller, Ryohsuke Mitsudome, Taekjin LEE

0.52.0 (2026-06-30)

0.51.0 (2026-05-01)

• Merge remote-tracking branch 'origin/main' into tmp/bot/bump_version_base

• feat(autoware_ground_segmentation_cuda): cuda 12.0 build compatibility (#12182)

  • feat(autoware_ground_segmentation_cuda): CUDA 12.0+ build compatibility

  * feat: restore Turing arch ---------

• Contributors: Amadeusz Szymko, github-actions

0.50.0 (2026-02-14)

• chore: match all package versions

• Merge remote-tracking branch 'origin/main' into humble

• feat(autoware_ground_segmentation_cuda): update nvcc flags (#12047)

• fix(agnocast): build on jazzy, remove from ground_segmentation_cuda (#11960)

• fix(cuda_scan_ground_segmentation): add missing include (#11850)

• feat: gpu ground segmentation (#11371)

  • init cuda ground segmentation node
  • feat: add cuda_scan_ground_segmentation node
  • fix: launch
  • refactor
  • refactor
  • fix: atan error
  • fix: num output point host reference
  • pre-commit
  • feat: add ground pointcloud publish for debug
  • refactor
  • refactor
  • refactor
  • fix: param path
  • feat: add debug processing time
  • delete pointcloud preprocesor ground segmentation
  • typo
  • refactoring
  • refactoring
  • fix: variable type
  • refactoring
  • typo
  • style(pre-commit): autofix
  • revert launch change
  • docs
  • style(pre-commit): autofix
  • revise variable type
  • fix: avoid very small d_radius
  • docs: add README
  • style(pre-commit): autofix
  • docs
  • refactor
  • style(pre-commit): autofix
  • docs: add schema
  • fix(recheck): classifiied local pointcloud index bug
  • fix: launch
  • remove vehicle info depend
  • Adding GPU ground segmentation
  • Finished coding
  • Fixed
  • Fixed
  • Fixed all. Debug code remains.
  • Remove debug code.
  • style(pre-commit): autofix

File truncated at 100 lines see the full file