点云配准

基于 ICP 的局部点云配准实验性 API：将源点云对齐到目标，并可选择使用初始位姿猜测。位于 Zivid::Experimental::Toolbox 中，可能会在未事先通知的情况下发生变化。

        %%{init: {'themeVariables': {'fontSize': '18px'}, 'flowchart': {'nodeSpacing': 30, 'rankSpacing': 35}}}%%
flowchart LR
  SourcePC["PointCloud (source)"]
  TargetPC["PointCloud (target)"]

  registrationOp(["localPointCloudRegistration()"])
  transformOp(["transform()"])
  residualOp(["residual()"])

  LocalPointCloudRegistrationResult
  Matrix4x4
  Residual["residual (mm)"]

  SourcePC -.-> registrationOp --> LocalPointCloudRegistrationResult
  TargetPC -.-> registrationOp
  LocalPointCloudRegistrationResult --> transformOp --> Matrix4x4
  LocalPointCloudRegistrationResult --> residualOp --> Residual

  classDef zividClass fill:#4A8FA4,stroke:#34323D,color:#FFFFFF
  classDef api fill:#91D2C8,stroke:#4A8FA4,color:#000000
  class SourcePC,TargetPC,LocalPointCloudRegistrationResult,Matrix4x4,Residual zividClass
  class registrationOp,transformOp,residualOp api

localPointCloudRegistration

使用 ICP 将源点云对齐到目标点云。当点云之间存在明显偏移时，提供可选的初始变换猜测可以提高收敛速度和准确度。

C++

LocalPointCloudRegistrationResult localPointCloudRegistration(const UnorganizedPointCloud &target, const UnorganizedPointCloud &source, const LocalPointCloudRegistrationParameters &params, const Pose &initialTransform = Pose{Matrix4x4::identity()})

Compute alignment transform between two point clouds.

Given a source point cloud and a target point cloud, this function attempts to compute the transform that must be applied to the source in order to align it with the target. This can be used to create a "stitched" unorganized point cloud of an object by combining data collected from different camera angles.

This function takes an argument initialTransform which is used as a starting-point for the computation of the transform that best aligns source with target. This initial guess is usually found from e.g. reference markers or robot capture pose, and this function is then used to refine the alignment. If the overlap of source and target is already quite good, one can pass the identity matrix as initialTransform.

The returned transform represents the total transform needed to align source with target, i.e. it includes both initialTransform and the refinement found by the algorithm.

Performance is very dependent on the number of points in either point cloud. To improve performance, voxel downsample one or both point clouds before passing them into this function. The resulting alignment transform can then be applied to the non-downsampled point clouds to still obtain a dense result.

Performance is also very dependent on MaxCorrespondenceDistance. To improve performance, try reducing this value. However, keep the value larger than the typical point-to-point distance in the point clouds, and larger than the expected translation error in the initial guess.

参数:

target -- The point cloud to align with
source -- The point cloud to be aligned with target
params -- Parameters for the registration process and its convergence criteria
initialTransform -- Initial guess applied to source point cloud before refinement

返回:

Instance of LocalPointCloudRegistrationResult

C#

class PointCloudRegistration

Point cloud registration functionality.

Public Static Functions

static LocalPointCloudRegistrationResult ^ LocalPointCloudRegistration (Zivid::NET::UnorganizedPointCloud^ target, Zivid::NET::UnorganizedPointCloud^ source, Zivid::NET::Experimental::LocalPointCloudRegistrationParameters^ parameters, Zivid::NET::Calibration::Pose^ initialTransform)

Compute alignment transform between two point clouds.

Given a source point cloud and a target point cloud, this function attempts to compute the transform that must be applied to the source in order to align it with the target. This can be used to create a "stitched" unorganized point cloud of an object by combining data collected from different camera angles.

This function takes an argument initialTransform which is used as a starting-point for the computation of the transform that best aligns source with target. This initial guess is usually found from e.g. reference markers or robot capture pose, and this function is then used to refine the alignment. If the overlap of source and target is already quite good, one can pass the identity matrix as initialTransform.

The returned transform represents the total transform needed to align source with target, i.e. it includes both initialTransform and the refinement found by the algorithm.

Performance is very dependent on the number of points in either point cloud. To improve performance, voxel downsample one or both point clouds before passing them into this function. The resulting alignment transform can then be applied to the non-downsampled point clouds to still obtain a dense result.

Performance is also very dependent on MaxCorrespondenceDistance. To improve performance, try reducing this value. However, keep the value larger than the typical point-to-point distance in the point clouds, and larger than the expected translation error in the initial guess.

参数:

target -- The point cloud to align with
source -- The point cloud to be aligned with target
parameters -- Parameters for the registration process and its convergence criteria
initialTransform -- Initial guess applied to source point cloud before refinement

返回:

Instance of LocalPointCloudRegistrationResult

Python

local_point_cloud_registration( target: UnorganizedPointCloud, source: UnorganizedPointCloud, parameters: LocalPointCloudRegistrationParameters | None = None, initial_transform: Pose | Matrix4x4 | ndarray | None = None, ) → LocalPointCloudRegistrationResult

Compute alignment transform between two point clouds.

Given a source point cloud and a target point cloud, this function attempts to compute the transform that must be applied to the source in order to align it with the target. This can be used to create a "stitched" unorganized point cloud of an object by combining data collected from different camera angles.

This function takes an argument initial_transform which is used as a starting-point for the computation of the transform that best aligns source with target. This initial guess is usually found from e.g. reference markers or robot capture pose, and this function is then used to refine the alignment. If the overlap of source and target is already quite good, one can pass the identity pose as initial_transform.

The returned transform represents the total transform needed to align source with target, i.e. it includes both initial_transform and the refinement found by the algorithm.

Performance is very dependent on the number of points in either point cloud. To improve performance, voxel downsample one or both point clouds before passing them into this function. The resulting alignment transform can then be applied to the non-downsampled point clouds to still obtain a dense result.

Performance is also very dependent on MaxCorrespondenceDistance. To improve performance, try reducing this value. However, keep the value larger than the typical point-to-point distance in the point clouds, and larger than the expected translation error in the initial guess.

参数:

target -- The target UnorganizedPointCloud to align with
source -- The source UnorganizedPointCloud to be aligned with target
parameters -- An optional LocalPointCloudRegistrationParameters instance. If not provided, default parameters are used.
initial_transform -- Optional transform (Pose/Matrix4x4/ndarray) be used as a starting point for the algorithm. If not provided, the identity transform is used.

返回:

A LocalPointCloudRegistrationResult instance

class LocalPointCloudRegistrationParameters( max_correspondence_distance: float | int = 10.0, max_iteration_count: int = 200, convergence_criteria: LocalPointCloudRegistrationParameters | None = None, )

Input parameters for controlling the local point cloud registration process.

Initialization

class ConvergenceCriteria( rmse_diff_threshold: float | int = 0.0001, source_coverage_diff_threshold: float | int = 0.0001, )

Criteria for ending iteration early. Iteration will end when all criteria are satisfied.

Initialization

property rmse_diff_threshold: Stop iteration when the absolute change in root mean square error (RMSE) from previous iteration is less than this threshold. RMSE is defined as the square root of the average squared distance between corresponding points in the source and target point clouds.

property source_coverage_diff_threshold: Stop iteration when the absolute change in coverage from previous iteration is less than this threshold. Coverage is defined as the fraction of points in the source point cloud that found a correspondence in the target point cloud.

property max_correspondence_distance: The radius around each source point in which the algorithm will look for neighbors in the target point cloud. Should be larger than the typical distance between points in the point clouds. This parameter may have a significant performance impact.

property max_iteration_count: End iteration when reaching this number, even if convergence criteria are not yet satisfied.

property convergence_criteria: Criteria for ending iteration early. Iteration will end when all criteria are satisfied.

classmethod load(file_name): Construct LocalPointCloudRegistrationParameters by loading from file

save(file_name): Save to the given file

classmethod from_serialized(value)

Construct a new LocalPointCloudRegistrationParameters instance from a previously serialized string

LocalPointCloudRegistrationResult

点云配准的结果。提供估计的 4×4 变换矩阵以及逐点重叠残差。

C++

class LocalPointCloudRegistrationResult

The result of a call to localPointCloudRegistration()

Public Functions

Pose transform() const: The transform that must be applied to the source point cloud for it to align with the target point cloud.

bool converged() const: A boolean indicating whether the convergence criteria were satisfied before reaching the iteration limit.

float sourceCoverage() const: The fraction of points in the source point cloud that has a correspondence in the target point cloud after transformation.

float rootMeanSquareError() const: The root mean squared distance between corresponding points in the source and target point cloud after transformation.

std::string toString() const: Get string representation of the result.

C#

class LocalPointCloudRegistrationResult

The result of a call to LocalPointCloudRegistration()

Public Functions

LocalPointCloudRegistrationResult(const Zivid::Experimental::Toolbox::LocalPointCloudRegistrationResult &nativeResult): Constructs a LocalPointCloudRegistrationResult from native result.

System::String ^ ToString () override: Get string representation of this object.

Properties

Zivid::NET::Calibration::Pose^ Transform: The transform that must be applied to the source point cloud for it to align with the target point cloud.

bool Converged: A boolean indicating whether the convergence criteria were satisfied before reaching the iteration limit.

float SourceCoverage: The fraction of points in the source point cloud that has a correspondence in the target point cloud after transformation.

float RootMeanSquareError: The root mean squared distance between corresponding points in the source and target point cloud after transformation.

Python

class LocalPointCloudRegistrationResult(impl)

The result of a call to local_point_cloud_registration.

Initialization

Initialize LocalPointCloudRegistrationResult wrapper.

This constructor is only used internally, and should not be called by the end-user.

参数:: impl -- Reference to internal/back-end instance.
抛出:: TypeError -- If argument does not match the expected internal class.

transform() → Pose

The transform that must be applied to the source point cloud for it to align with the target point cloud.

返回:: A rigid transform as a Pose instance

converged() → bool

A boolean indicating whether the convergence criteria were satisfied before reaching the iteration limit.

返回:: A boolean indicating convergence

source_coverage() → float

The fraction of points in the source that has a correspondence in the target after transformation.

返回:: A float in the range [0.0, 1.0]

root_mean_square_error() → float

The root mean squared distance between corresponding points in the source and target after transformation.

返回:: A float representing the root mean square error

另请参阅