研究目的
To propose a novel method for 3D point cloud registration that improves accuracy and computational efficiency by using convex hulls on projected points and refining with ICP or NDT algorithms.
研究成果
The proposed two-step registration method, combining coarse registration via convex hulls on projected points with refinement by ICP or NDT, is effective for improving accuracy and reducing computation time in 3D point cloud registration. CHACR ICP generally provides better results than CHACR NDT and standard ICP, especially for large transformations. Future work should address noise elimination, edge similarity testing in convex hulls, and parallelization for further efficiency gains.
研究不足
The method is sensitive to noise and outliers, requiring pre-processing steps for filtering. It assumes no noise or outliers in the data, which may not hold in real-world scenarios. The refinement with NDT does not consistently outperform standard NDT, and the approach may not handle all types of non-rigid transformations effectively.
1：Experimental Design and Method Selection:
The study employs a two-step registration strategy: coarse registration using convex hulls on projected points, followed by fine registration using ICP or NDT. PCA is used for plane determination, and convex hulls are matched to estimate initial transformations.
2：Sample Selection and Data Sources:
Real 3D point clouds (e.g., blade, bunny, dragon) are acquired from a 3D scanner repository [1], with synthetic transformations applied for testing.
3：List of Experimental Equipment and Materials:
A PC with 8 GB RAM and Intel Core i5 processor, running Windows 7 and Visual Studio 2010. Software includes Armadillo library for PCA, quickHull algorithm for convex hull computation, and implementations of ICP and NDT from a previous platform [5].
4：Software includes Armadillo library for PCA, quickHull algorithm for convex hull computation, and implementations of ICP and NDT from a previous platform [5].
Experimental Procedures and Operational Workflow:
4. Experimental Procedures and Operational Workflow: Steps include noise filtering, projection onto PCA-determined planes, convex hull extraction and matching, initial transformation estimation using Singular Value Decomposition, and refinement with ICP or NDT on projected points. Performance is evaluated based on standard deviation error and computation time.
5：Data Analysis Methods:
Standard deviation of Euclidean distances between corresponding points is used to measure accuracy. Computation time is recorded in milliseconds. Results are visualized using semi-logarithmic scales for error and time comparisons.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
