Research on Inter-Frame Feature Mismatch Removal Method of VSLAM in Dynamic Scenes

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Visual Simultaneous Localization and Mapping (VSLAM) is crucial for robot pose estimation.
• Feature point mismatches in VSLAM degrade localization and mapping accuracy, especially in dynamic environments.
• Existing methods struggle with accurate feature matching in scenes with moving objects.

Purpose of the Study:

• To propose and evaluate a novel method for removing inter-frame feature mismatches in dynamic scenes for visual SLAM.
• To enhance the robustness and accuracy of mobile robot self-localization and mapping.

Main Methods:

• Introduced Grid-based Motion Statistics (GMS) for rapid coarse screening of mismatched features.
• Developed Adaptive Error Threshold RANSAC (ATRANSAC) for accurate mismatch removal based on internal matching rate.
• Combined GMS and ATRANSAC into the GMS-ATRANSAC method for robust performance in dynamic and static scenes.

Main Results:

• GMS-ATRANSAC effectively removes feature mismatches on moving objects.
• Achieved average error reduction of 29.4% (vs. RANSAC) and 32.9% (vs. GMS-RANSAC).
• Reduced error variance by 63.9% (vs. RANSAC) and 58.0% (vs. GMS-RANSAC), with processing time reductions of 78.3% and 38%, respectively.
• Validated effectiveness in ORB-SLAM2 initialization and ORB-SLAM3 tracking threads.

Conclusions:

• The proposed GMS-ATRANSAC method significantly improves feature mismatch removal in dynamic visual SLAM.
• This leads to more accurate robot localization and mapping, outperforming RANSAC and GMS-RANSAC.
• The algorithm enhances the reliability of VSLAM systems like ORB-SLAM2 and ORB-SLAM3.