Study on Dynamic-Point Filters in Visual SLAM Systems
Research evaluates the impact of dynamic-point filters on feature-based visual SLAM.
This research investigates the effects of dynamic-point filters in feature-based visual SLAM systems, particularly focusing on the challenges posed by low-texture environments.
Research Overview
The study presents a controlled experiment using synthetic indoor sequences to analyze how varying surface texture and scene dynamics influence SLAM performance.
Methodology
Different filtering methods were compared, including ORB-SLAM2 without filtering, an optical-flow filter, and a multi-view depth-consistency filter. The evaluation metrics included trajectory error and tracking completeness.
Findings
Results indicated that filtering was beneficial primarily in dynamic scenes, but in low-texture scenarios, it could hinder tracking completeness. Notably, the GEOM filter was found to discard more static keypoints compared to FLOW.