Integrating voxel mapping with deep network-based point-line feature fusion for robust SLAM

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Visual SLAM systems often struggle with feature loss and map consistency in challenging environments like low-texture, low-light, and unstructured scenes.
• Existing methods face limitations in maintaining robustness and accuracy under these adverse conditions.

Purpose of the Study:

• To propose an improved visual SLAM system addressing feature loss and map consistency issues.
• To enhance matching robustness and 3D reconstruction quality in challenging visual SLAM scenarios.

Main Methods:

• A novel approach combining custom voxel mapping with sparse SLAM for improved feature matching and reconstruction.
• Integration of a depth map neural network for effective fusion of point and line features.
• Evaluation using public datasets and real-world unstructured scenes.

Main Results:

• Significant improvement in joint feature matching efficiency by leveraging complementary point and line features.
• Enhanced 3D reconstruction quality and matching robustness in low-texture regions.
• Demonstrated superior localization accuracy and environmental adaptability in challenging scenarios.

Conclusions:

• The proposed visual SLAM system offers a substantial advancement for reliable operation in difficult real-world conditions.
• The fusion of point and line features, combined with advanced mapping techniques, proves effective in overcoming common SLAM limitations.
• This work lays the groundwork for more robust and dependable SLAM applications.