PDC-Net: Robust point cloud registration using deep cyclic neural network combined with PCA.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Data Processing

Background:

• Accurate and efficient registration of 3D point clouds is crucial for various applications.
• Traditional methods like Iterative Closest Point (ICP) can be slow and prone to local optima.
• Deep learning approaches offer potential for improved registration performance.

Purpose of the Study:

• To propose a novel, robust deep learning-based point cloud registration method.
• To enhance both the precision and speed of the registration process.
• To overcome limitations of existing registration algorithms, particularly regarding initial alignment and local optima.

Main Methods:

• Developed PDC-Net, a two-stage deep learning framework for point cloud registration.
• Stage 1: Principal Component Analysis (PCA)-based adjustment network for rapid initial pose estimation.
• Stage 2: Iterative neural network employing an inverse compositional algorithm for precise final transformation.

Main Results:

• PDC-Net demonstrated competitive registration accuracy compared to state-of-the-art methods on the ModelNet40 dataset.
• The method showed robustness to varying initial point cloud phase differences, avoiding local optima.
• Achieved improved registration speed compared to traditional and some learning-based approaches.

Conclusions:

• PDC-Net offers a robust and efficient solution for point cloud registration.
• The deep learning architecture effectively handles initial alignment and iterative refinement.
• The proposed method enhances registration reliability and performance in challenging scenarios.