Study on Robustness of ORB-SLAM Based Outlier Elimination in Bronchoscope Tracking - RANSAC + EPnP for Outlier Detection

摘要

In this paper, we investigate the robustness of outlier removal in bronchoscope posture tracking based on ORB-SLAM. Our method uses an image-based posture tracking method called ORB-SLAM to track the posture of bronchoscope. ORB feature points are extracted from the image in each frame of the real bronchoscope video, the position and orientation of the bronchoscope are estimated from the positional relationship of the feature points between two consecutive frames, 3D structure of the bronchi is simultaneously reconstructed. However, due to the deformation of the bronchus, outlier matches may exist in matched feature points, which may result in larger posture estimation errors. In this paper, candidate matches are selected using RANSAC in order to reduce mismatch in calculation of corresponding points. Furthermore, after estimating the posture using the EPnP algorithm, candidate matches are selected by calculating reprojection errors. Experimental results showed that the method proposed in four cases can track an average of 224 frames more than the previous tracking system. The average tracking rate of the proposed method was 76.5%.