autoware_map_height_fitter

This library fits the given point with the ground of the point cloud map. The map loading operation is switched by the parameter enable_partial_load of the node specified by map_loader_name. The node using this library must use multi thread executor.

Parameters

{{ json_to_markdown(“map/autoware_map_height_fitter/schema/map_height_fitter.schema.json”) }}

Topic subscription

Service client

Changelog for package autoware_map_height_fitter

1.9.0 (2026-06-24)

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

• perf(autoware_map_height_fitter): extract testable ground-height kernel and use a cached KdTree (#1107)

  * perf(autoware_map_height_fitter): extract testable ground-height kernel and use a cached KdTree Extract the pure ground-height search out of the Node/TF-bound pimpl Impl into free functions in a new internal header so the core algorithm becomes unit-testable without ROS:

  • get_ground_height_from_pointcloud(cloud[, kdtree], x, y, fallback_z)
  • get_ground_height_from_vector_map(map, x, y, fallback_z)

  - build_pointcloud_xy_kdtree(cloud) Impl::get_ground_height now delegates to these. The two full O(N) linear passes over the point cloud per fit() request are replaced with a pcl::KdTreeFLANN built once when the cloud is set (on_pcd_map / get_partial_point_cloud_map) and reused: nearestKSearch finds the closest point and radiusSearch restricts the lowest-z scan to a local candidate set, dropping per-call cost from O(N) to ~O(log N + k). Behavior is preserved: the KdTree indexes the horizontal (z = 0) projection so the 2D distance metric is unchanged, and the membership test is recomputed in double precision with the original strict '< (sqrt(min_dist2) + 1.0)\^2' comparison over the candidate set, so the selected point set (and thus the returned height) is identical to the original two-pass scan. The vector-map nearest-point lookup and all non-finite / empty / missing-map fallbacks are unchanged. Also fold in two cheap, behavior-preserving micro-optimizations flagged in the audit: radius2 is computed as r*r instead of std::pow(..., 2.0), and the partial-map merge loop reserves the total data size before concatenating grids. Add a gtest suite (test/test_map_height_fitter_kernel.cpp) covering the empty-cloud and single-point fallbacks, lowest-z-within-radius selection, the horizontal-only distance metric, the prebuilt-vs-convenience overload equivalence, the vector-map nearest height and its non-finite/empty fallbacks, and a 300-trial randomized characterization test asserting the new kernel matches the original two-pass scan exactly. The package previously had zero tests. Refs: autowarefoundation/autoware_core#1096

  * perf(autoware_map_height_fitter): preserve double-precision nearest search The KdTree path searched entirely in float: both the query and the indexed coordinates were rounded to float, and min_dist2 was taken from the single index returned by nearestKSearch. At large map coordinates (MGRS/UTM-derived) or for near-tied points the float nearest can differ from the true double-precision nearest, changing the radius and the selected height, so the documented bit-for-bit behavior preservation did not hold (demonstrably ~1 in 2000 at coordinates up to 2e6). Recompute min_dist2 in double precision over a margin-enlarged candidate set that is a guaranteed superset of the double-nearest point, and size both the nearest and the height radiusSearch margins from the coordinate magnitude (2 * coord_max * 2\^-23 + 1e-3) instead of a fixed 1e-3, so every point the strict double-precision test accepts is in the candidate set regardless of coordinate scale. The double-precision filter then reproduces the original membership exactly. Add a large-coordinate / near-tie characterization test (5000 trials, centers up to 2e6) that asserts bit-equality against the original two-pass scan; it fails on the previous float-only logic and passes now. Refs: autowarefoundation/autoware_core#1096

  * test(autoware_map_height_fitter): pin non-finite kernel inputs and guard NaN query (#67) Add edge-case kernel tests for non-finite cloud coords, query, and z against the in-test double-precision oracle; guard a non-finite query so the KdTree path returns fallback_z instead of tripping a FLANN assert, preserving the pre-PR scan behavior. Refs: autowarefoundation/autoware_core#1096

  • ci(autoware_map_height_fitter): add cspell:ignore for distance var names

  * refactor(autoware_map_height_fitter): move internal kernel header to src/ Address review: map_height_fitter_kernel.hpp is only used within this package (nothing outside includes it), so move it out of the public include/ tree into src/ as a non-installed internal header. Switch its include guard to the bare-filename form required for src/ headers, point the three includers at the relative path, and let the gtest reach it via target_include_directories(PRIVATE src). ---------

• Contributors: Yutaka Kondo, github-actions

1.8.0 (2026-05-01)

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

• refactor(autoware_core): add USE_SCOPED_HEADER_INSTALL_DIR to map

File truncated at 100 lines see the full file