PlaneFusion: Real-Time Indoor Scene Reconstruction With Planar Prior

Area of Science:

• Computer Vision
• Robotics
• 3D Reconstruction

Background:

• Real-time dense Simultaneous Localization and Mapping (SLAM) is crucial for 3D scene reconstruction using RGB or RGB-D sensors.
• Indoor environments present challenges like large low-texture regions that degrade reconstruction quality.
• Planar priors offer a potential solution to enhance reconstruction in these challenging areas.

Purpose of the Study:

• To fully explore the utility of planar priors within a dense SLAM pipeline.
• To develop a novel, efficient plane detection and segmentation method for real-time SLAM.
• To introduce a memory-efficient plane-based map representation for dense SLAM.

Main Methods:

• A novel plane detection and segmentation algorithm operating at 200 Hz on a GPU.
• Efficient construction of global plane constraints integrated into each frame's camera pose estimation.
• A new plane-based map representation designed to minimize memory footprint while preserving geometric details.

Main Results:

• The proposed system achieves superior 3D reconstruction results using planar information.
• Real-time performance exceeding 30 frames per second (fps) is maintained.
• Effective handling of low-texture regions within planar areas, a common issue in indoor SLAM.

Conclusions:

• Integrating planar priors significantly improves real-time dense SLAM performance and reconstruction quality.
• The developed methods offer both speed and memory efficiency for 3D scene reconstruction.
• This approach effectively addresses the challenge of low-texture regions in indoor environments.