Research on a Density-Based Clustering Method for Eliminating Inter-Frame Feature Mismatches in Visual SLAM Under...

Area of Science:

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

Background:

• Visual SLAM systems rely on static feature points for accurate localization and map construction.
• Dynamic feature points introduce errors in motion pose estimation, compromising SLAM system accuracy and robustness.

Purpose of the Study:

• To develop a method for eliminating feature mismatches in dynamic scenes within visual SLAM.
• To enhance the accuracy and robustness of visual SLAM systems by addressing challenges posed by dynamic environments.

Main Methods:

• A spatial clustering-based RANSAC method is proposed to differentiate and remove dynamic feature points, creating a high-quality static dataset.
• The refined dataset is then processed using RANSAC to efficiently eliminate local mismatches by fitting geometric models.
• The method, termed DSSAC-RANSAC, is integrated into ORB-SLAM2 and ORB-SLAM3 for validation.

Main Results:

• The DSSAC-RANSAC method significantly reduces average reprojection error by up to 58.5% compared to traditional RANSAC and GMS-RANSAC.
• Reprojection error variance is decreased by up to 65.2%, indicating improved stability.
• Processing time is reduced by up to 69.4%, enhancing computational efficiency.

Conclusions:

• The proposed DSSAC-RANSAC algorithm effectively eliminates feature mismatches in dynamic visual SLAM scenes.
• The method demonstrably improves accuracy, reduces error variance, and speeds up processing, enhancing overall SLAM robustness.
• Integration into established SLAM frameworks like ORB-SLAM2 and ORB-SLAM3 confirms its practical applicability and effectiveness.