Low-Latency Visual-Based High-Quality 3-D Reconstruction Using Point Cloud Optimization

摘要

In recent years, 3-D reconstruction has been widely used in robot pose estimation, mine exploration, building of digital twins, and other fields. Visual-based reconstruction methods can restore the color of objects while building a 3-D point cloud map, which makes the map more intuitive. However, although the current visual-based methods have good real-time performance and high frequency, it is slightly inadequate in some scenes that require high accuracy but have no obvious demand for update frequency, such as robot welding scenes and professional service robot positioning. Therefore, a new visual-based pose estimation and 3-D reconstruction method based on image feature extraction and point cloud recognition was proposed in this research, which can improve the accuracy of visual-based pose estimation methods by 3-D point cloud matching. To verify the effectiveness of the method, the current commonly used algorithms are selected for comparison in outdoor binocular camera scenes, indoor red green blue-depth (RGB-D) camera scenes, and actual welding robot scenes. The results showed that the method proposed in this research significantly improved the reconstruction accuracy while ensuring low-latency performance. Based on the high-precision environmental awareness and rapid response positioning, the application of this method will make welding manufacturing and maintenance more automatic and intelligent.