2HR-Net VSLAM: Robust visual SLAM based on dual high-reliability feature matching in dynamic environments

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Visual Simultaneous Localization and Mapping (VSLAM) is crucial for mobile robot navigation.
• Existing feature-based VSLAM systems struggle with feature reliability and dynamic interference in complex environments.

Purpose of the Study:

• To develop a dynamic adaptive VSLAM system that improves localization accuracy in challenging, dynamic environments.
• To enhance feature matching robustness and reduce the impact of dynamic elements on pose estimation.

Main Methods:

• Proposed a High-repeatability and High-reliability feature matching network (2HR-Net) for robust feature detection.
• Integrated K-Means clustering into L2-Net for high-repeatability and high-reliability feature point detection.
• Utilized a lightweight YOLOv8n model for real-time detection and removal of dynamic feature points.
• Developed a shared matching Siamese network with dual-branch feature fusion and similarity optimization for enhanced matching accuracy.

Main Results:

• The 2HR feature detection achieved approximately 70% repeatability in dynamic scenarios, surpassing traditional methods (below 40%).
• Demonstrated a significant reduction (approx. 90%) in Absolute Trajectory Error (ATE) RMSE and S.D. compared to ORB-SLAM3.
• Validated superior performance in feature repeatability, matching accuracy, and localization precision on the TUM dataset.

Conclusions:

• The proposed dynamic adaptive VSLAM system offers a robust solution for autonomous navigation in complex, dynamic environments.
• The 2HR-Net and integrated YOLOv8n model effectively address limitations of traditional VSLAM approaches.
• Achieved state-of-the-art performance in localization accuracy and stability, outperforming ORB-SLAM3.