Multi-view Point Cloud Registration based on Evolutionary Multitasking with Bi-Channel Knowledge Sharing Mechanism

Yue Wu, Yibo Liu, Maoguo Gong, Hao Li, Zedong Tang, Qiguang Miao, Wenping Ma

Abstract

This paper introduces MTPCR, a framework that reformulates multi-view 3D point cloud registration as a multi-task optimization problem. By leveraging a novel "bi-channel" knowledge-sharing mechanism, the system simultaneously optimizes local alignment accuracy and global consistency. This approach effectively prevents the "drift" errors and local optima common in traditional 3D reconstruction, ensuring a high-fidelity unified coordinate system for complex 3D scans.

Introduction

When scanning a 3D object or environment, sensors capture multiple "point clouds" from various perspectives. Reconstructing the complete scene requires aligning (registering) these clouds into a single coordinate system. Traditional methods struggle with two main hurdles:

  1. Local Optima: Pairwise registration is often "sticky," where algorithms get trapped in incorrect alignments due to noise or low overlap.
  2. Error Accumulation: Sequential alignment (e.g., A to B, then B to C) leads to cumulative errors. Without considering global geometry, a visible "gap" often appears when closing the loop, a phenomenon known as a lack of global consistency.

The MTPCR Framework

The authors propose MTPCR (Multi-tasking Point Cloud Registration), which utilizes Evolutionary Multitasking (EMT). Instead of solving one alignment at a time, MTPCR treats the entire set of registrations as a collection of related tasks solved concurrently.

Bi-Channel Knowledge Sharing

The core of MTPCR is its dual-channel mechanism for transferring information between tasks:

  • Intra-task Sharing: For every difficult registration task, the algorithm creates an "aiding task"—a simplified version of the problem. Genetic material (solutions) flows between the aiding task and the original task to accelerate convergence and provide a "warm start."
  • Inter-task Sharing: This exploits the topological relationships between different pairs of clouds. If a set of scans forms a loop, the algorithm enforces a "loop closure constraint" (where the product of transformations equals the Identity matrix). Best solutions from one task are encoded and transferred to others to pull them out of local optima.
Graph representation of multi-view registration tasks showing loop structures used for inter-task sharing.

Robust Fitness Function

To evaluate the quality of an alignment, MTPCR uses a robust loss function that maximizes the "consensus set"—the number of points that overlap within a specific distance. The system employs an adaptive threshold ($\epsilon$) calculated from the average distance between nearest neighbors, making it self-tuning for different data scales and sensor types.

Experimental Results

The researchers evaluated MTPCR on several benchmarks, including the Stanford 3D Scanning Repository (Bunny, Dragon), the Aim@Shape Project, and the Augmented ICL-NUIM Dataset (real-world indoor scenes).

The multi-view registration process: individual scans from different angles (left) are aligned into a globally consistent model (right).

Key Insights

  • Superior Accuracy: MTPCR outperformed traditional methods like Trimmed ICP (TrICP) and K-Means registration, particularly in rotation error reduction.
  • Stability in Noise: The consensus-based fitness function makes MTPCR significantly more resilient to Gaussian noise and outliers than standard mean-squared error approaches.
  • Convergence Speed: The "Intra-task" channel effectively guides the search process, making it more efficient than standard evolutionary multitasking models.

Conclusion and Future Work

MTPCR demonstrates that treating registration as a multi-task problem acts as a powerful regularizer, forcing the algorithm to find solutions that satisfy the entire scene's geometry rather than just local pairs. While more computationally intensive than local gradient-descent methods, its robustness and global consistency offer significant advantages for complex reconstruction tasks. Future work will explore using machine learning to further refine the knowledge transfer process between tasks.

Summary updated

Abstract

Multi-view point cloud registration is fundamental in 3D reconstruction. Since there are close connections between point clouds captured from different viewpoints, registration performance can be enhanced if these connections be harnessed properly. Therefore, this paper models the registration problem as multi-task optimization, and proposes a novel bi-channel knowledge sharing mechanism for effective and efficient problem solving. The modeling of multi-view point cloud registration as multi-task optimization are twofold. By simultaneously considering the local accuracy of two point clouds as well as the global consistency posed by all the point clouds involved, a fitness function with an adaptive threshold is derived. Also a framework of the co-evolutionary search process is defined for the concurrent optimization of multiple fitness functions belonging to related tasks. To enhance solution quality and convergence speed, the proposed bi-channel knowledge sharing mechanism plays its role. The intra-task knowledge sharing introduces aiding tasks that are much simpler to solve, and useful information is shared across aiding tasks and the original tasks, accelerating the search process. The inter-task knowledge sharing explores commonalities buried among the original tasks, aiming to prevent tasks from getting stuck to local optima. Comprehensive experiments conducted on model object as well as scene point clouds show the efficacy of the proposed method.
BibTeX
@article{wu2022multi,
  author = {Yue Wu and Yibo Liu and Maoguo Gong and Hao Li and Zedong Tang and Qiguang Miao and Wenping Ma},
  title = {{Multi-view Point Cloud Registration based on Evolutionary Multitasking with Bi-Channel Knowledge Sharing Mechanism}},
  journal = {IEEE Transactions on Emerging Topics in Computational Intelligence},
  year = {2022},
  doi = {10.1109/tetci.2022.3205384},
}
Figure