A Dynamic Graph Convolutional Point Cloud Registration Method Combining Global Attention
- Resource Type
- Conference
- Authors
- Lv, Zhengkun; Yi, Kaixiang; Zhou, Wenju
- Source
- 2023 China Automation Congress (CAC) Automation Congress (CAC), 2023 China. :2126-2131 Nov, 2023
- Subject
- Aerospace
Communication, Networking and Broadcast Technologies
Components, Circuits, Devices and Systems
Computing and Processing
Power, Energy and Industry Applications
Robotics and Control Systems
Signal Processing and Analysis
Transportation
Point cloud compression
Heuristic algorithms
Feature extraction
Robustness
Convolutional neural networks
Iterative methods
Task analysis
point cloud registration
DGCNN
attention mechanism
key point extraction
- Language
- ISSN
- 2688-0938
In order to solve the problems of low efficiency and high requirements on the initial position of aligned point clouds encountered by the traditional ICP (Iterative Closest Point) registration algorithm in the registration of large point cloud models, this paper proposes a point cloud registration algorithm based on the global attention mechanism and dynamic graph convolutional network. The algorithm enhances the spatial connection between point cloud pairs by introducing a global attention mechanism to enhance the extraction of key features in point clouds and integrating this mechanism in DGCNN (Dynamic Graph Convolutional Neural Network) feature extraction networks. In addition, the algorithm uses a similarity-based soft connection instead of the traditional hard connection for point cloud point pair matching, resulting in faster processing speed and higher registration accuracy. Experimental results show that the algorithm exhibits higher robustness and registration accuracy in large point cloud models, demonstrating the effectiveness of the global attention mechanism in the point cloud registration task.