Gerald Liu∗ and Chao Shen∗∗
Visual SLAM (vSLAM), monocular SLAM, frame interpolation,feature filtering, ORB-SLAM3, robotics
Feature-based visual SLAM (vSLAM) systems, such as ORB- SLAM3, rely on robust feature matching. However, it can be disrupted by low feature quality and tracking inconsistencies caused by fast rotations. While this issue can be partially mitigated by generating intermediate frames through video frame interpolation models, these methods introduce visual artifacts, which can degrade the reliability of interpolated features, causing errors in pose estimation. This paper presents two multi-scale entropy- based feature filtering techniques to minimise the side effects of artifacts and improve pose estimation accuracy in interpolation- assisted vSLAM. The first method employs a multi-scale spatial entropy analysis within interpolated frames to reject keypoints from low-information regions while the second enforces frame-to- frame temporal consistency by comparing entropy distributions to a reliable baseline from the last non-interpolated frame to identify and filter out features from areas likely distorted by interpolation artifacts. Our experiments on KITTI and EuRoC demonstrate that the proposed filtering methods successfully mitigate these artifacts and outperform the baseline vSLAM system in many scenarios. Moreover, our methods consistently improve trajectory smoothness and yield significant improvements in trajectory accuracy. Our findings emphasise the important role of targeted artifact mitigation for integrating interpolation techniques in modern SLAM pipelines.
Important Links:
Go Back
