Abstract
Point Cloud Registration (PCR) is a critical task in robotics and 3D reconstruction, often plagued by local optima in high-dimensional search spaces. This paper presents Evolutionary Multi-Task Registration with Solution Space Cutting (EMTR-SSC), a novel framework that leverages evolutionary multitasking to improve registration robustness. By concurrently solving a full 6D registration task and a simplified rotation-only task using Translation Invariant Measurements (TIMs), the algorithm facilitates knowledge transfer to guide the search out of mathematical traps. Experimental results across diverse datasets demonstrate that EMTR-SSC significantly outperforms existing evolutionary and deep learning methods in success rate and precision while maintaining computational efficiency through a sparse-to-dense strategy.
Introduction
Point Cloud Registration (PCR) involves finding the optimal transformation to align two 3D datasets into a common coordinate frame. While foundational for autonomous systems and 3D modeling, PCR remains challenging due to the non-convex nature of the objective function. Traditional local search methods like Iterative Closest Point (ICP) are highly sensitive to initial poses and frequently converge to local optima. Evolutionary Computation (EC) offers a more global search perspective but often suffers from high computational costs and susceptibility to noise or partial overlaps.
Methodology
The EMTR-SSC approach treats registration as a multitasking problem. It defines two distinct but related tasks:
- Task $\alpha$ (Main Task): A 6D search space encompassing 3D rotation and 3D translation using the original point cloud data.
- Task $\beta$ (Auxiliary Task): A reduced-dimension task focusing solely on rotation. This "solution space cutting" utilizes Translation Invariant Measurements (TIMs), which remain constant regardless of the object's translation, simplifying the search landscape.
The framework incorporates several key innovations:
- Knowledge Transfer & Complement: As Task $\beta$ identifies promising rotation parameters, this information is transferred to Task $\alpha$. To bridge the gap between the rotation-only Task $\beta$ and the 6D Task $\alpha$, a Knowledge Complement strategy estimates the missing translation components using weighted least squares.
- M-estimator Chamfer Distance (MCD): A robust fitness function that mitigates the impact of outliers and partial overlaps by ignoring points without clear correspondences.
- Sparse-to-Dense Strategy: To optimize performance, the algorithm initially operates on downsampled point clouds, transitioning to full-density data only during the final refinement phase.
Results and Discussion
The effectiveness of EMTR-SSC was validated across 15 datasets, including object-scale (Stanford, U3M, FGR), synthetic (ModelNet40), and scene-scale (RESSO) data.
Key findings include:
- Robustness: EMTR-SSC achieved an average success ratio of 87.7%, a substantial improvement over other evolutionary multitasking methods which typically ranged between 17% and 55%.
- Accuracy: The method consistently yielded lower rotation and translation errors compared to traditional algorithms (FGR, TrICP) and deep learning models (RPM-Net).
- Efficiency: The sparse-to-dense mechanism effectively halved the average runtime from approximately 15 seconds to 7.39 seconds.
The "scout" role of Task $\beta$ is pivotal. By navigating a simpler landscape, it avoids the local traps that hinder Task $\alpha$. The synergy between these tasks allows the population to explore the search space more effectively, pulling the main task toward the global optimum.
Conclusion
EMTR-SSC addresses the local optimum problem in PCR through a strategic multitasking approach. By combining solution space cutting with robust fitness evaluation and efficient data handling, it provides a powerful tool for complex 3D alignment tasks. Future work will focus on developing adaptive parameter mechanisms to further enhance the algorithm's versatility across varying point cloud complexities.
Summary updated