Enhanced RGB-D SLAM through orthogonal plane constraints and point-line-plane collaborative optimization

Area of Science:

• Computer Vision
• Robotics
• Simultaneous Localization and Mapping (SLAM)

Background:

• Visual localization in complex indoor environments is challenging due to feature degradation and cumulative errors.
• Existing SLAM frameworks often struggle with global drift and lack robustness in diverse conditions.

Purpose of the Study:

• To propose PLPM-SLAM, a novel RGB-D SLAM framework enhancing robustness and accuracy through orthogonal Manhattan plane constraints and point-line-plane joint optimization.
• To mitigate global drift by jointly decoupling rotation and translation using three mutually orthogonal planes.
• To improve performance in unstructured and low-texture environments using virtual plane construction and vanishing-point-guided optimization.

Main Methods:

• Integration of orthogonal Manhattan plane constraints for joint rotation and translation decoupling.
• Development of a virtual plane construction strategy for incomplete Manhattan structures.
• Application of homogeneous and heterogeneous geometric constraints in tracking and optimization.
• Implementation of a vanishing-point-guided joint optimization model for unstructured environments.

Main Results:

• PLPM-SLAM demonstrates superior performance over ORB-SLAM3 on public (TUM, ICL-NUIM) and real-world datasets.
• Achieved significant Root Mean Square Error (RMSE) reductions, up to 82.77% on public datasets and 92.16% on real-world data.
• Consistently enhanced accuracy and robustness in both structured and low-texture indoor environments.

Conclusions:

• PLPM-SLAM offers a robust and accurate solution for visual localization in challenging indoor environments.
• The proposed framework effectively addresses global drift and improves geometric consistency.
• PLPM-SLAM represents a significant advancement in RGB-D SLAM technology, outperforming state-of-the-art methods.