Skip to content

kiss_icp_gpu_core.md

Latest commit

 

History

History
74 lines (57 loc) · 2.57 KB

kiss_icp_gpu_core.md

File metadata and controls

74 lines (57 loc) · 2.57 KB
Raw

Reusable GPU KISS-ICP Core

cudarobotics_kiss_icp_gpu is the streaming LiDAR-odometry core used by the gpu_kiss_icp demo and intended for the CudaNav ROS 2 component. It has no OpenCV dependency and never consumes ground truth.

API

Include:

#include <cudarobotics/kiss_icp_gpu.hpp>

Create one persistent odometry instance, choose the coordinate-system anchor explicitly, and submit tightly packed x, y, z points:

cudarobotics::KissIcpConfig config;
config.map_voxel_size = 0.5f;
config.scan_voxel_size = 0.5f;
config.nn_backend = cudarobotics::KissIcpNnBackend::Voxel;

cudarobotics::KissIcpOdometry odometry(config);
cudarobotics::KissIcpPose initial_pose;  // identity by default
odometry.reset(initial_pose);

for (const std::vector<float>& xyz : scans) {
    auto frame = odometry.register_scan(xyz);
    publish_odometry(frame.pose);
}

The first scan initializes the voxel map at initial_pose. Every later call uses the previous estimate as the prediction, computes adaptive-threshold point-to-plane ICP against the local map, and inserts the registered scan.

Contract

  • Input is finite sensor-frame XYZ in metres, with exactly three floats per point.
  • Output pose is T_world_sensor; the caller chooses world in reset().
  • A KissIcpOdometry instance owns persistent CUDA buffers and is not copyable.
  • Scan and local-map capacities are explicit. Overflow raises an exception; points are never silently truncated.
  • The voxel NN hash capacity must be a power of two and at least the configured map capacity.
  • map_snapshot() returns the current first-observation voxel map in world coordinates.
  • KissIcpFrameResult exposes input/sample/map counts and per-frame ICP diagnostics. timing() exposes accumulated map upload, normal estimation, and index-build time.

Configuration can be checked before CUDA allocation with validate_kiss_icp_config().

Verification

cmake --build build --target cudarobotics_kiss_icp_gpu \
  kiss_icp_gpu_streaming_smoke gpu_kiss_icp -j"$(nproc)"
ctest --test-dir build -R 'kiss_icp_gpu_streaming_smoke|gpu_kiss_icp_gate' \
  --output-on-failure

The streaming smoke verifies explicit-pose initialization, stationary-scan registration, malformed input rejection, and reset semantics. The existing 64-frame benchmark remains the accuracy and correspondence-performance gate.

CudaNav boundary

The ROS 2 lifecycle component will own this class. PointCloud2 decoding, timestamp checks, TF publication, diagnostics, and lifecycle recovery remain ROS responsibilities and are intentionally outside this CUDA core.