Point Cloud Registration Using Hierarchical Equivariant 9DoF Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing point cloud registration methods struggle to accurately align non-uniformly scaled source and target point clouds, which are common in robotics tasks such as category-level object pose estimation and indoor reconstruction, due to the increased complexity of estimating a 9DoF transformation.
Innovation Solution
A hierarchical equivariant graph neural network (HEGN) is proposed to directly estimate a 9DoF transformation by leveraging equivariance to rotation, translation, and scaling, using centroid subtraction, equivariant feature extraction, and hierarchical graph neural networks to determine 3DoF rotation, translation, and scaling components.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional 6DoF transformation methods are used for point cloud registration, then the registration process is computationally efficient and simple, but registration accuracy deteriorates when dealing with non-uniformly scaled point clouds
Solution Approach 1:
The patent segments the 9DoF transformation estimation into three independent 3DoF components (rotation, translation, scaling). Each component is estimated separately using dedicated neural network branches, which simplifies the overall complex estimation task while maintaining high accuracy for non-uniformly scaled point clouds
Solution Approach 2:
The patent extends the transformation space from conventional 6DoF to 9DoF by adding the scaling dimension. This dimensional extension allows the model to handle non-uniform scaling transformations, improving registration accuracy for scaled point clouds while using equivariant neural networks to manage the increased complexity
2Measurement precision
If 9DoF transformation estimation is implemented to handle non-uniform scaling, then registration accuracy for scaled point clouds improves, but computational complexity increases
Solution Approach 1:
The transformation estimation is segmented into three parallel 3DoF estimation tasks (rotation, translation, scaling) processed by separate neural network branches. This segmentation enables independent optimization of each component and parallel computation, improving computational efficiency while maintaining the capability to handle 9DoF transformations
Solution Approach 2:
The patent uses equivariant neural networks that inherently preserve transformation properties (rotation equivariance, translation equivariance, scaling equivariance). This parameter-preserving approach reduces the number of learnable parameters and computational operations needed, improving efficiency while accurately estimating 9DoF transformations
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention relates to a method for determining a transformation for registering a target point cloud to a source point cloud wherein for each shifted point cloud a respective equivariant node feature set is generated for respective graph hierarchy levels m of a hierarchical graph comprising M graph hierarchy levels, by an equivariant feature extraction and invariant node pooling procedure, wherein the hierarchical graph is representing the respective point cloud; a hierarchical global equivariant feature set of the respective graph hierarchy level m of the respective graph is generated by a global pooling procedure depending on the equivariant node feature set of the respective graph hierarchy level; a further global equivariant feature set of the respective graph is generated by a hierarchical aggregation procedure depending on the respective hierarchical global equivariant feature sets of the graphs; the transformation is determined depending on the global equivariant feature sets.